perf: bit-exact batch fast paths + streaming-first benchmark docs (#202)

## Summary - Dedicated batch fast paths for **EMA, RSI, Bollinger, MACD and ATR** (used by the Python bindings): one allocation filled in a single pass, warmup encoded as `NaN`, no per-element `Option` or input re-validation. Each is **bit-for-bit equal** to replaying `update` — SMA/Bollinger keep the drift-reseed cadence, the EMA-family keep the seed division and `mul_add` recurrences. Adds the `BatchNanExt` extension trait. - **Cross-library benchmark refresh**: `compare_libraries.py` reports the median across timing rounds (`--rounds` / `--streaming-rounds`), gains `--skip-batch` / `--skip-streaming`, and runs every peer through the streaming arena (recompute for batch-only libraries). `wickra-bench` drives the batch fast paths against `kand`. - **README** benchmark section reordered streaming-first (the order-of-magnitude result), with measured TA-Lib/tulipy/pandas-ta numbers in place of the CI-only placeholders. ## Impact - Python batch ~2× faster on EMA/RSI/MACD/ATR; streaming path unchanged. - The `batch == streaming` equivalence stays bit-exact. ## Verification - `cargo fmt` · `cargo clippy --workspace --all-targets --all-features -- -D warnings` (clean) - `cargo test --workspace --all-features` — 3782 unit + 420 doc tests pass - Python `pytest` — streaming-vs-batch, known-values, input-validation, smoke pass ## Notes - Node/WASM bindings keep their existing batch; the fast paths are Python-only for now.
2026-06-08 00:17:58 +02:00
parent e97c3389fe
commit 05fe7ffa90
14 changed files with 1400 additions and 296 deletions
@@ -0,0 +1,96 @@
+# Benchmarks
+
+Read these as **relative** speedups on identical input — absolute µs depend on
+CPU, memory clock and OS scheduler, not a universal contract. **Streaming is the
+headline**: it is where Wickra's design pays off and where the gap is measured in
+orders of magnitude, not percent. The batch numbers come second and are shown
+honestly — the leanest crates edge Wickra out on the simple recurrences, and that
+is a deliberate trade for warmup/NaN semantics, not a ceiling.
+
+- **Reproduced on:** Windows 11 Pro 26200, AMD Ryzen 9 9950X, 64 GB DDR5,
+  Rust 1.92 (release: `lto = "fat"`, `codegen-units = 1`), Python 3.12.
+- **Reproduce yourself:**
+  - Rust core vs Rust crates: `cargo bench -p wickra-bench`
+  - Python vs Python libs: `pip install -e bindings/python[bench]` then
+    `python -m benchmarks.compare_libraries` (auto-detects installed peers).
+
+## 1. Streaming — the structural win
+
+Live trading feeds one tick at a time. Wickra updates every indicator in **O(1)**;
+batch-only libraries (TA-Lib, tulipy, finta, pandas-ta) have no incremental API
+and must recompute the whole history on every tick. Only `talipp` (Python) and
+`ta-rs` / `yata` (Rust) carry real per-tick state. This is the gap the library
+was built to expose.
+
+**Python — per-tick latency** (seed 5 000 bars, then feed ticks one at a time):
+
+| Indicator        | **★&nbsp;Wickra** | talipp           | TA-Lib (recompute)    |
+|------------------|------------------:|------------------|-----------------------|
+| SMA(20)          | **0.063 µs ★**    | 0.59 µs (9×)     | 204 µs (3 300×)       |
+| EMA(20)          | **0.060 µs ★**    | 0.72 µs (12×)    | 212 µs (3 500×)       |
+| RSI(14)          | **0.065 µs ★**    | 1.06 µs (16×)    | 230 µs (3 600×)       |
+| MACD(12, 26, 9)  | **0.078 µs ★**    | 4.22 µs (54×)    | 245 µs (3 100×)       |
+| Bollinger(20, 2) | **0.088 µs ★**    | 5.15 µs (58×)    | 229 µs (2 600×)       |
+
+Against the only other incremental Python peer Wickra is **9–58× faster**;
+against the recompute-on-every-tick libraries it is **2 600–14 000× faster**
+(`finta` RSI hits 14 000×). tulipy / pandas-ta land in the same recompute band
+as TA-Lib.
+
+**Rust — per-tick latency** (whole 50 000-bar series, lower = faster):
+
+| Indicator        | **★&nbsp;Wickra** | kand | ta-rs | yata |
+|------------------|------------------:|-----:|------:|-----:|
+| SMA(20)          | 50                | 38   | 47    | 38   |
+| EMA(20)          | 154               | 69   | 56    | 69   |
+| RSI(14)          | 164               | 216  | 74    | —    |
+| MACD(12, 26, 9)  | 275               | 143  | 66    | —    |
+| Bollinger(20, 2) | **128 ★**         | 248  | 168   | —    |
+| ATR(14)          | 152               | 166  | 61    | —    |
+
+`ta-rs` hands back a bare `f64` from the first tick with no warmup and no
+validation; it leads several rows by giving those guarantees up. Against `kand`,
+Wickra wins streaming RSI, Bollinger and ATR. `yata` exposes only SMA/EMA as
+raw-value methods, so its other rows are omitted rather than faked.
+
+## 2. Batch — competitive, not the headline
+
+Whole series in one call. Here hand-tuned C (`tulipy`, TA-Lib) and the leanest
+Rust crate (`kand`) win the simple recurrences — Wickra trades a few µs per pass
+for the `None`-warmup, NaN-safety and bit-exact `batch == streaming` guarantees
+none of them keep. It still wins several rows outright and beats the rest of the
+field everywhere.
+
+**Python** (20 000-bar pass, µs/op, lower = faster):
+
+| Indicator        | Wickra   | TA-Lib | tulipy | pandas-ta |
+|------------------|---------:|-------:|-------:|----------:|
+| SMA(20)          | 22.7     | **15.4** | 15.9 | 33.7      |
+| EMA(20)          | 30.8     | **30.3** | 31.1 | 48.8      |
+| RSI(14)          | 58.9     | 72.5   | **38.5** | 94.8    |
+| MACD(12, 26, 9)  | 71.7     | 99.1   | **33.5** | 207.6   |
+| Bollinger(20, 2) | 84.9     | 65.7   | **32.3** | 336.4   |
+| ATR(14)          | 52.0     | 79.4   | **31.9** | —        |
+
+Wickra beats TA-Lib on RSI, MACD and ATR and the whole Python field on every
+row; tulipy's SIMD C stays ahead on the heavier indicators.
+
+**Rust** (50 000-bar pass, µs, lower = faster). Only Wickra and `kand` expose a
+batch API; `ta-rs` and `yata` are streaming-only:
+
+| Indicator        | **★&nbsp;Wickra** | kand   |
+|------------------|------------------:|-------:|
+| SMA(20)          | 53                | **41** |
+| EMA(20)          | 111               | **71** |
+| RSI(14)          | **221 ★**         | 259    |
+| MACD(12, 26, 9)  | 533               | **327** |
+| Bollinger(20, 2) | **404 ★**         | 460    |
+| ATR(14)          | **122 ★**         | 169    |
+
+Run the suite yourself:
+
+```bash
+cargo bench -p wickra-bench            # Rust core vs kand / ta-rs / yata
+pip install -e bindings/python[bench]  # Python peers
+python -m benchmarks.compare_libraries
+```
@@ -11,6 +11,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Andrews Pitchfork** — median line and two parallels projected from the last three swing pivots (`ANDREWS_PITCHFORK`).
 - **Murrey Math Lines** — T. H. Murrey's eighths grid over the recent trading range, each level acting as support/resistance (`MURREY_MATH_LINES`).
 - **Central Pivot Range** — the classic pivot flanked by two central levels gauging the day's expected character (`CENTRAL_PIVOT_RANGE`).
+- **Faster scalar batch paths** — `Ema`, `Rsi`, `BollingerBands`, `MacdIndicator` and `Atr` gained dedicated batch fast paths (used by the Python bindings) that strip per-element `Option`/validation overhead and the intermediate `Vec<Option<_>>` allocation, while staying *bit-for-bit* equal to replaying `update` (including the SMA/Bollinger drift-reseed). Python batch is ~2× faster on EMA/RSI/MACD/ATR; streaming is unchanged.
+- **Cross-library benchmark refresh** — `benchmarks/compare_libraries.py` now measures the median across timing rounds (`--rounds` / `--streaming-rounds`), adds `--skip-batch` / `--skip-streaming`, and drives every peer through the streaming arena (recompute for batch-only libraries). `wickra-bench` compares the batch fast paths against `kand`.

 ## [0.6.5] - 2026-06-07
 - **Autocorrelation Periodogram** — Ehlers autocorrelation periodogram: dominant cycle period estimate (`AUTOCORRPGRAM`).
@@ -58,19 +58,44 @@ Full documentation lives at **[docs.wickra.org](https://docs.wickra.org)**:
  [TA-Lib migration](https://docs.wickra.org/TA-Lib-Migration),
  [FAQ](https://docs.wickra.org/FAQ).

-## Why Wickra exists
+## Why Wickra

-Wickra started as a personal itch. The existing TA libraries never quite fit the
-projects I was building, so I decided to build one from the ground up — partly to
-learn, partly because I genuinely enjoy taking something that already exists and
-trying to do it differently (and, ideally, better). It's open source because the
-useful version of that itch is the one other people can build on too.
+Most TA libraries are fast, *or* multi-language, *or* broad. Wickra refuses to
+pick. It's the streaming-first engine built for the workload the others treat as
+an afterthought — **live, tick-by-tick data** — without giving up the breadth of
+a full batch library, and without making you reimplement your indicators four
+times to get there.

-Plenty of TA libraries are fast. Each one forces a trade-off Wickra does not:
+- **The biggest streaming-native catalogue, period.** 467 indicators across 24
+  families — candlesticks, harmonic & chart patterns, market profile, market
+  breadth, Renko/Kagi/Point&Figure bars, Ehlers DSP cycles, risk/performance
+  metrics — every single one updating in **O(1) per tick**. TA-Lib ships ~150 and
+  none of them stream.
+- **One Rust core, four first-class targets.** Native **Python · Node.js ·
+  WebAssembly · Rust** — identical math, identical results, zero per-language
+  reimplementation and zero GIL bottleneck.
+- **Correct by construction, not by hope.** Every `update` validates its input,
+  runs a real warmup, and returns an `Option` so a single bad tick can't silently
+  poison state. `batch == streaming` is **bit-exact, fuzzed and 100 %-line-covered
+  for all 467 indicators**.
+- **Orders of magnitude faster where it counts.** In streaming Wickra is **9–58×**
+  faster than the only other incremental peer and **thousands of times** faster
+  than recompute-on-every-tick libraries. On batch it wins several rows outright
+  and trades the simple recurrences (SMA, EMA, MACD) for its guarantees — and
+  the losses are shown, not hidden.
+- **Install in one line, anywhere.** `pip install wickra` / `npm install wickra` —
+  precompiled wheels and binaries, **no C toolchain, none of TA-Lib's setup pain**.
+  macOS · Linux · Windows.
+- **Batteries included.** Indicator chaining, a streaming OHLCV CSV reader, and a
+  live Binance kline feed ship in the box.
+- **Truly permissive.** **MIT OR Apache-2.0** — drop it straight into commercial
+  and closed-source work.
+
+Every other library forces one of those compromises. Wickra doesn't:

 | Library          | Install     | Streaming   | Languages                   | Indicators | Active |
 |------------------|-------------|-------------|-----------------------------|-----------:|--------|
-| **★&nbsp;Wickra**| **clean**   | **yes, O(1)** | **Python · Node · WASM · Rust** | **423** | **yes** |
+| **★&nbsp;Wickra**| **clean**   | **yes, O(1)** | **Python · Node · WASM · Rust** | **467** | **yes** |
 | kand             | clean       | yes         | Python · WASM · Rust        |       ~60  | yes    |
 | ta-rs            | clean       | yes         | Rust only                   |       ~30  | stale  |
 | yata             | clean       | partial     | Rust only                   |       ~35  | yes    |
@@ -79,112 +104,27 @@ Plenty of TA libraries are fast. Each one forces a trade-off Wickra does not:
 | finta            | clean       | no          | Python                      |       ~80  | stale  |
 | talipp           | clean       | yes         | Python                      |       ~40  | yes    |

-Wickra's edge is **breadth with reach**: 467 indicators that all update in O(1)
-per tick and ship natively to Python, Node.js, WebAssembly and Rust from a
-single engine.
+Broad, multi-language, streaming-native **and** honest about its trade-offs — at
+the same time. That's the combination no one else ships.

-**On speed — and why Wickra isn't the fastest.** It deliberately isn't. The
-leaner Rust crates (kand, ta-rs) win several of the micro-benchmarks below, and
-those losses are shown rather than hidden. The gap is a *choice*, not a ceiling:
-every `update` validates its input, runs a real warmup before it emits a value,
-and returns an `Option` so a single bad tick can't silently poison the state.
-ta-rs, by contrast, hands back a bare `f64` from the first tick with no
-validation. If Wickra threw all of that away — raw `f64` out, no checks, no
-warmup contract — it would match or beat the leanest crate on every row. It
-keeps the guarantees instead, and still wins RSI, Bollinger and ATR against kand.
-What no other library matches is the *combination*: catalogue size, native O(1)
-streaming, NaN-safety, and four first-class language targets at once.
+## Why Wickra exists
+
+Wickra started as a personal itch. The existing TA libraries never quite fit the
+projects I was building, so I decided to build one from the ground up — partly to
+learn, partly because I genuinely enjoy taking something that already exists and
+trying to do it differently (and, ideally, better). It's open source because the
+useful version of that itch is the one other people can build on too.

 ## Benchmarks

-Three comparisons, split by layer and mode. Read them as **relative** speedups
-on identical input — absolute µs depend on CPU, memory clock and OS scheduler,
-not a universal contract.
+Wickra updates every indicator in **O(1)** per tick. In **streaming** — the
+workload it is built for — it is **9–58× faster** than the only other incremental
+peer and **thousands of times** faster than recompute-on-every-tick libraries.
+**Batch** is competitive: it wins several rows outright and trades a few µs
+elsewhere for `None`-warmup, NaN-safety and bit-exact `batch == streaming`.

- **Reproduced on:** Windows 11 Pro 26200, AMD Ryzen 9 9950X, 64 GB DDR5,
-  Rust 1.92 (release: `lto = "fat"`, `codegen-units = 1`), Python 3.12.
- **Reproduce yourself:**
-  - Rust core vs Rust crates: `cargo bench -p wickra-bench`
-  - Python vs Python libs: `pip install -e bindings/python[bench]` then
-    `python -m benchmarks.compare_libraries` (auto-detects installed peers).
-
-### 1. Rust core vs the other Rust TA crates
-
-Like-for-like, no language-binding overhead, over a 50 000-bar series (µs for
-the whole series, lower = faster). This is the honest engine comparison —
-Wickra wins some and loses some, and both are shown.
-
-**Streaming** (one value fed per `update`):
-
-| Indicator        | **★&nbsp;Wickra** | kand | ta-rs | yata |
-|------------------|------------------:|-----:|------:|-----:|
-| SMA(20)          | 50                | 38   | 47    | 38   |
-| EMA(20)          | 154               | 69   | 56    | 69   |
-| RSI(14)          | 164               | 216  | 74    | —    |
-| MACD(12, 26, 9)  | 275               | 143  | 66    | —    |
-| Bollinger(20, 2) | **128 ★**         | 248  | 168   | —    |
-| ATR(14)          | 152               | 166  | 61    | —    |
-
-**Batch** (whole series at once). Only Wickra and kand expose a batch API;
-ta-rs and yata are streaming-only.
-
-| Indicator        | **★&nbsp;Wickra** | kand |
-|------------------|------------------:|-----:|
-| SMA(20)          | 82                | 42   |
-| EMA(20)          | 159               | 74   |
-| RSI(14)          | **253 ★**         | 274  |
-| MACD(12, 26, 9)  | 681               | 283  |
-| Bollinger(20, 2) | **445 ★**         | 462  |
-| ATR(14)          | 175               | 173  |
-
-ta-rs is the per-indicator speed champion on almost every row — it returns a
-bare `f64` with no warmup state and no input validation, trading away the
-`None`-warmup and NaN-safety semantics Wickra keeps. Against kand, Wickra wins
-streaming RSI, Bollinger and ATR (and batch RSI + Bollinger); Bollinger is the
-one row where Wickra is the outright fastest of all four. The leaner crates
-still win the pure recurrences (EMA, MACD) and SMA. yata exposes only SMA/EMA as
-raw-value methods, so its other rows are omitted rather than faked.
-
-### 2. Python vs the Python TA ecosystem — batch
-
-Full pass over a 20 000-bar series, µs/op (lower = faster). **★** per row.
-
-| Indicator        | **★&nbsp;Wickra** | finta               | TA-Lib | tulipy |
-|------------------|------------------:|---------------------|--------|--------|
-| SMA(20)          | **59.6 ★**        | 354.2 (5.9× slower) | ⧗      | ⧗      |
-| EMA(20)          | **88.4 ★**        | 309.3 (3.5× slower) | ⧗      | ⧗      |
-| RSI(14)          | **77.3 ★**        | 1 283 (16.6× slower)| ⧗      | ⧗      |
-| MACD(12, 26, 9)  | **116.4 ★**       | 529.5 (4.6× slower) | ⧗      | ⧗      |
-| Bollinger(20, 2) | **146.0 ★**       | 1 246 (8.5× slower) | ⧗      | ⧗      |
-| ATR(14)          | **135.8 ★**       | 3 812 (28× slower)  | ⧗      | ⧗      |
-
-> ⧗ = published by the CI Linux job. TA-Lib and tulipy ship C extensions that
-> don't build cleanly on every desktop, so their canonical numbers come from the
-> `cross-library-bench` workflow rather than this local table. pandas-ta needs
-> Python ≥ 3.12 and isn't in the 3.11 CI matrix. The script auto-detects
-> whichever peers are installed in your environment.
-
-### 3. Python — streaming (per-tick latency)
-
-Seed 5 000 bars, then feed ticks one at a time. talipp is the only Python peer
-with a true incremental API; batch-only libraries like TA-Lib must recompute the
-entire history on every tick — Wickra updates in O(1).
-
-| Indicator        | **★&nbsp;Wickra (per tick)** | talipp (per tick)       |
-|------------------|------------------------------:|-------------------------|
-| SMA(20)          | **0.067 µs ★**                | 0.63 µs (9.4× slower)   |
-| EMA(20)          | **0.051 µs ★**                | 0.63 µs (12.2× slower)  |
-| RSI(14)          | **0.053 µs ★**                | 1.00 µs (19.1× slower)  |
-| MACD(12, 26, 9)  | **0.071 µs ★**                | 3.64 µs (51.5× slower)  |
-| Bollinger(20, 2) | **0.085 µs ★**                | 4.87 µs (57.2× slower)  |
-
-Run the suite yourself:
-
-```bash
-cargo bench -p wickra-bench            # Rust core vs kand / ta-rs / yata
-pip install -e bindings/python[bench]  # Python peers
-python -m benchmarks.compare_libraries
-```
+Full tables (Rust + Python, streaming + batch) and how to reproduce them live in
+**[BENCHMARKS.md](BENCHMARKS.md)**.

 ## Indicators

@@ -72,13 +72,23 @@ class Sample:
        return (self.seconds / self.iterations) * 1_000_000


-def time_call(fn: Callable[[], None], iterations: int) -> float:
-    """Time ``fn`` over ``iterations`` calls, returning total wall seconds."""
+def time_call(fn: Callable[[], None], iterations: int, rounds: int = 5) -> float:
+    """Time ``fn`` over ``iterations`` calls per round, across ``rounds`` rounds.
+
+    Returns the *median* round's wall seconds for one round of ``iterations``
+    calls. Taking the median across several rounds damps the OS scheduling and
+    GC jitter that a single timing pass would otherwise bake into the result,
+    so the per-iteration figure is stable run-to-run. Callers keep dividing the
+    return value by ``iterations``.
+    """
    fn()  # one warmup call to populate caches
-    start = time.perf_counter()
-    for _ in range(iterations):
-        fn()
-    return time.perf_counter() - start
+    rounds_s: List[float] = []
+    for _ in range(rounds):
+        start = time.perf_counter()
+        for _ in range(iterations):
+            fn()
+        rounds_s.append(time.perf_counter() - start)
+    return statistics.median(rounds_s)


 def gen_prices(n: int, seed: int = 0xC0FFEE) -> np.ndarray:
@@ -457,6 +467,161 @@ def talipp_bollinger_streaming(seed: np.ndarray, live: np.ndarray) -> Optional[C
    return run


+# Recompute streaming peers: batch-only libraries have no incremental API, so
+# the only honest way to drive them tick-by-tick is to re-run the full batch
+# over the grown history on every new price. These runners expose exactly that
+# cost — the gap Wickra's O(1) update closes.
+
+
+def _talib_recompute_streaming(seed, live, fn):
+    def run() -> None:
+        history = list(seed)
+        for p in live:
+            history.append(float(p))
+            fn(np.asarray(history))
+
+    return run
+
+
+def _pandas_ta_recompute_streaming(seed, live, fn):
+    def run() -> None:
+        history = list(seed)
+        for p in live:
+            history.append(float(p))
+            fn(PD.Series(history))
+
+    return run
+
+
+def _tulipy_recompute_streaming(seed, live, fn):
+    def run() -> None:
+        history = list(seed)
+        for p in live:
+            history.append(float(p))
+            fn(np.asarray(history, dtype=np.float64))
+
+    return run
+
+
+def _finta_recompute_streaming(seed, live, fn):
+    def run() -> None:
+        history = list(seed)
+        for p in live:
+            history.append(float(p))
+            arr = np.asarray(history)
+            fn(PD.DataFrame({"open": arr, "high": arr, "low": arr, "close": arr, "volume": np.ones_like(arr)}))
+
+    return run
+
+
+def talib_sma_streaming(seed, live):
+    if TALIB is None:
+        return None
+    return _talib_recompute_streaming(seed, live, lambda a: TALIB.SMA(a, timeperiod=20))
+
+
+def pandas_ta_sma_streaming(seed, live):
+    if PANDAS_TA is None or PD is None:
+        return None
+    return _pandas_ta_recompute_streaming(seed, live, lambda s: PANDAS_TA.sma(s, length=20))
+
+
+def tulipy_sma_streaming(seed, live):
+    if TULIPY is None:
+        return None
+    return _tulipy_recompute_streaming(seed, live, lambda a: TULIPY.sma(a, 20))
+
+
+def finta_sma_streaming(seed, live):
+    if FINTA is None or PD is None:
+        return None
+    return _finta_recompute_streaming(seed, live, lambda df: FINTA.TA.SMA(df, period=20))
+
+
+def talib_ema_streaming(seed, live):
+    if TALIB is None:
+        return None
+    return _talib_recompute_streaming(seed, live, lambda a: TALIB.EMA(a, timeperiod=20))
+
+
+def pandas_ta_ema_streaming(seed, live):
+    if PANDAS_TA is None or PD is None:
+        return None
+    return _pandas_ta_recompute_streaming(seed, live, lambda s: PANDAS_TA.ema(s, length=20))
+
+
+def tulipy_ema_streaming(seed, live):
+    if TULIPY is None:
+        return None
+    return _tulipy_recompute_streaming(seed, live, lambda a: TULIPY.ema(a, 20))
+
+
+def finta_ema_streaming(seed, live):
+    if FINTA is None or PD is None:
+        return None
+    return _finta_recompute_streaming(seed, live, lambda df: FINTA.TA.EMA(df, period=20))
+
+
+def tulipy_rsi_streaming(seed, live):
+    if TULIPY is None:
+        return None
+    return _tulipy_recompute_streaming(seed, live, lambda a: TULIPY.rsi(a, 14))
+
+
+def finta_rsi_streaming(seed, live):
+    if FINTA is None or PD is None:
+        return None
+    return _finta_recompute_streaming(seed, live, lambda df: FINTA.TA.RSI(df, period=14))
+
+
+def talib_macd_streaming(seed, live):
+    if TALIB is None:
+        return None
+    return _talib_recompute_streaming(seed, live, lambda a: TALIB.MACD(a))
+
+
+def pandas_ta_macd_streaming(seed, live):
+    if PANDAS_TA is None or PD is None:
+        return None
+    return _pandas_ta_recompute_streaming(seed, live, lambda s: PANDAS_TA.macd(s))
+
+
+def tulipy_macd_streaming(seed, live):
+    if TULIPY is None:
+        return None
+    return _tulipy_recompute_streaming(seed, live, lambda a: TULIPY.macd(a, 12, 26, 9))
+
+
+def finta_macd_streaming(seed, live):
+    if FINTA is None or PD is None:
+        return None
+    return _finta_recompute_streaming(seed, live, lambda df: FINTA.TA.MACD(df))
+
+
+def talib_bollinger_streaming(seed, live):
+    if TALIB is None:
+        return None
+    return _talib_recompute_streaming(seed, live, lambda a: TALIB.BBANDS(a, timeperiod=20, nbdevup=2, nbdevdn=2))
+
+
+def pandas_ta_bollinger_streaming(seed, live):
+    if PANDAS_TA is None or PD is None:
+        return None
+    return _pandas_ta_recompute_streaming(seed, live, lambda s: PANDAS_TA.bbands(s, length=20, std=2.0))
+
+
+def tulipy_bollinger_streaming(seed, live):
+    if TULIPY is None:
+        return None
+    return _tulipy_recompute_streaming(seed, live, lambda a: TULIPY.bbands(a, 20, 2.0))
+
+
+def finta_bollinger_streaming(seed, live):
+    if FINTA is None or PD is None:
+        return None
+    return _finta_recompute_streaming(seed, live, lambda df: FINTA.TA.BBANDS(df, period=20, std_multiplier=2.0))
+
+
 # --------------------------------------------------------------------------- #
 # Runner
 # --------------------------------------------------------------------------- #
@@ -519,36 +684,54 @@ STREAMING_INDICATORS = [
    ("SMA(20)", [
        ("Wickra", wickra_sma_streaming),
        ("talipp", talipp_sma_streaming),
+        ("TA-Lib", talib_sma_streaming),
+        ("pandas-ta", pandas_ta_sma_streaming),
+        ("tulipy", tulipy_sma_streaming),
+        ("finta", finta_sma_streaming),
    ]),
    ("EMA(20)", [
        ("Wickra", wickra_ema_streaming),
        ("talipp", talipp_ema_streaming),
+        ("TA-Lib", talib_ema_streaming),
+        ("pandas-ta", pandas_ta_ema_streaming),
+        ("tulipy", tulipy_ema_streaming),
+        ("finta", finta_ema_streaming),
    ]),
    ("RSI(14)", [
        ("Wickra", wickra_rsi_streaming),
+        ("talipp", talipp_rsi_streaming),
        ("TA-Lib", talib_rsi_streaming),
        ("pandas-ta", pandas_ta_rsi_streaming),
-        ("talipp", talipp_rsi_streaming),
+        ("tulipy", tulipy_rsi_streaming),
+        ("finta", finta_rsi_streaming),
    ]),
    ("MACD(12, 26, 9)", [
        ("Wickra", wickra_macd_streaming),
        ("talipp", talipp_macd_streaming),
+        ("TA-Lib", talib_macd_streaming),
+        ("pandas-ta", pandas_ta_macd_streaming),
+        ("tulipy", tulipy_macd_streaming),
+        ("finta", finta_macd_streaming),
    ]),
    ("Bollinger(20, 2.0)", [
        ("Wickra", wickra_bollinger_streaming),
        ("talipp", talipp_bollinger_streaming),
+        ("TA-Lib", talib_bollinger_streaming),
+        ("pandas-ta", pandas_ta_bollinger_streaming),
+        ("tulipy", tulipy_bollinger_streaming),
+        ("finta", finta_bollinger_streaming),
    ]),
 ]


-def run_batch(prices: np.ndarray, iterations: int) -> List[Sample]:
+def run_batch(prices: np.ndarray, iterations: int, rounds: int) -> List[Sample]:
    out: List[Sample] = []
    for indicator_name, libs in BATCH_INDICATORS:
        for lib_name, factory in libs:
            runner = factory(prices)
            if runner is None:
                continue
-            secs = time_call(runner, iterations)
+            secs = time_call(runner, iterations, rounds)
            out.append(Sample(lib_name, indicator_name, "batch", secs, iterations))
    return out

@@ -558,6 +741,7 @@ def run_ohlc(
    low: np.ndarray,
    close: np.ndarray,
    iterations: int,
+    rounds: int,
 ) -> List[Sample]:
    out: List[Sample] = []
    for indicator_name, libs in OHLC_INDICATORS:
@@ -565,12 +749,12 @@ def run_ohlc(
            runner = factory(high, low, close)
            if runner is None:
                continue
-            secs = time_call(runner, iterations)
+            secs = time_call(runner, iterations, rounds)
            out.append(Sample(lib_name, indicator_name, "batch", secs, iterations))
    return out


-def run_streaming(prices: np.ndarray, streaming_window: int, iterations: int) -> List[Sample]:
+def run_streaming(prices: np.ndarray, streaming_window: int, iterations: int, rounds: int) -> List[Sample]:
    out: List[Sample] = []
    seed = prices[:streaming_window]
    live = prices[streaming_window:]
@@ -581,7 +765,7 @@ def run_streaming(prices: np.ndarray, streaming_window: int, iterations: int) ->
            runner = factory(seed, live)
            if runner is None:
                continue
-            secs = time_call(runner, iterations)
+            secs = time_call(runner, iterations, rounds)
            sample = Sample(lib_name, indicator_name, "streaming", secs, iterations)
            sample.iterations = iterations * len(live)  # per-tick normalization
            out.append(sample)
@@ -629,6 +813,12 @@ def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0] if __doc__ else None)
    parser.add_argument("--size", type=int, default=20_000, help="number of prices")
    parser.add_argument("--iterations", type=int, default=20, help="batch repetitions per timing")
+    parser.add_argument(
+        "--rounds",
+        type=int,
+        default=5,
+        help="batch timing rounds; the median round is reported to damp jitter",
+    )
    parser.add_argument(
        "--streaming-window",
        type=int,
@@ -641,6 +831,14 @@ def parse_args() -> argparse.Namespace:
        default=3,
        help="repetitions of the streaming workload (each iteration replays all live ticks)",
    )
+    parser.add_argument(
+        "--streaming-rounds",
+        type=int,
+        default=2,
+        help="streaming timing rounds; the median round is reported",
+    )
+    parser.add_argument("--skip-batch", action="store_true", help="skip the batch tables")
+    parser.add_argument("--skip-streaming", action="store_true", help="skip the streaming tables")
    return parser.parse_args()


@@ -660,11 +858,14 @@ def main() -> None:
    print(f"Streaming window: {args.streaming_window} seed, {args.size - args.streaming_window} live")

    high, low, close, _ = gen_ohlc(args.size)
-    batch_rows = run_batch(prices, args.iterations)
-    ohlc_rows = run_ohlc(high, low, close, args.iterations)
-    streaming_rows = run_streaming(prices, args.streaming_window, args.streaming_iterations)
+    rows: List[Sample] = []
+    if not args.skip_batch:
+        rows += run_batch(prices, args.iterations, args.rounds)
+        rows += run_ohlc(high, low, close, args.iterations, args.rounds)
+    if not args.skip_streaming:
+        rows += run_streaming(prices, args.streaming_window, args.streaming_iterations, args.streaming_rounds)

-    print(render_table(batch_rows + ohlc_rows + streaming_rows))
+    print(render_table(rows))


 if __name__ == "__main__":
@@ -25,7 +25,7 @@

 use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
 use std::hint::black_box;
-use wickra::{Atr, BatchExt, BollingerBands, Candle, Ema, Indicator, MacdIndicator, Rsi, Sma};
+use wickra::{Atr, BollingerBands, Candle, Ema, Indicator, MacdIndicator, Rsi, Sma};
 use wickra_data::csv::CandleReader;
 use yata::prelude::Method;

@@ -82,7 +82,7 @@ fn sma_group(crit: &mut Criterion, closes: &[f64]) {
            |bencher, &series| {
                bencher.iter(|| {
                    let mut ind = Sma::new(SMA_PERIOD).unwrap();
-                    black_box(ind.batch(series));
+                    black_box(ind.batch_nan(series));
                });
            },
        );
@@ -170,7 +170,7 @@ fn ema_group(crit: &mut Criterion, closes: &[f64]) {
            |bencher, &series| {
                bencher.iter(|| {
                    let mut ind = Ema::new(EMA_PERIOD).unwrap();
-                    black_box(ind.batch(series));
+                    black_box(ind.batch_nan(series));
                });
            },
        );
@@ -253,7 +253,7 @@ fn rsi_group(crit: &mut Criterion, closes: &[f64]) {
            |bencher, &series| {
                bencher.iter(|| {
                    let mut ind = Rsi::new(RSI_PERIOD).unwrap();
-                    black_box(ind.batch(series));
+                    black_box(ind.batch_nan(series));
                });
            },
        );
@@ -352,7 +352,7 @@ fn macd_group(crit: &mut Criterion, closes: &[f64]) {
            |bencher, &series| {
                bencher.iter(|| {
                    let mut ind = MacdIndicator::classic();
-                    black_box(ind.batch(series));
+                    black_box(ind.batch_macd(series));
                });
            },
        );
@@ -478,7 +478,7 @@ fn bbands_group(crit: &mut Criterion, closes: &[f64]) {
            |bencher, &series| {
                bencher.iter(|| {
                    let mut ind = BollingerBands::new(BB_PERIOD, BB_DEV).unwrap();
-                    black_box(ind.batch(series));
+                    black_box(ind.batch_bands(series));
                });
            },
        );
@@ -604,9 +604,13 @@ fn atr_group(crit: &mut Criterion, candles: &[Candle]) {
            BenchmarkId::new("wickra/batch", len),
            &series,
            |bencher, &series| {
+                // Column extraction is outside the timed loop, mirroring kand's arm.
+                let high: Vec<f64> = series.iter().map(|candle| candle.high).collect();
+                let low: Vec<f64> = series.iter().map(|candle| candle.low).collect();
+                let close: Vec<f64> = series.iter().map(|candle| candle.close).collect();
                bencher.iter(|| {
                    let mut ind = Atr::new(ATR_PERIOD).unwrap();
-                    black_box(ind.batch(series));
+                    black_box(ind.batch_atr(&high, &low, &close));
                });
            },
        );
@@ -75,6 +75,67 @@ impl Atr {
            None
        }
    }
+
+    /// Vectorized batch over raw high/low/close columns: one `f64` per bar
+    /// (`NaN` during warmup). The caller guarantees the three slices are equal
+    /// length and finite with valid OHLC ordering (the binding validates once up
+    /// front); ATR only reads high, low and the previous close.
+    ///
+    /// For a fresh indicator long enough to seed (`n >= period`) it runs the
+    /// true-range seed once and then the bare Wilder recurrence in a tight loop —
+    /// no per-bar `Candle` construction/validation, no `Option`, identical
+    /// division at the seed and `mul_add` afterwards, so the result is
+    /// *bit-for-bit* equal to replaying `update` over the same candles. Shorter
+    /// or non-fresh inputs defer to an exact `update` replay.
+    pub fn batch_atr(&mut self, high: &[f64], low: &[f64], close: &[f64]) -> Vec<f64> {
+        let p = self.period;
+        let n = high.len();
+        if self.seeded || !self.seed_buf.is_empty() || self.prev_close.is_some() || n < p {
+            let mut out = vec![f64::NAN; n];
+            for i in 0..n {
+                let candle = Candle::new_unchecked(close[i], high[i], low[i], close[i], 0.0, 0);
+                if let Some(v) = self.update(candle) {
+                    out[i] = v;
+                }
+            }
+            return out;
+        }
+
+        // Warmup `[0, p-1)` is `NaN`; the first ATR is emitted at index `p - 1`.
+        let mut out = vec![f64::NAN; p - 1];
+        out.reserve(n - (p - 1));
+        // Seed: mean of the first `period` true ranges. TR₀ has no previous close.
+        let mut prev_close = close[0];
+        let mut sum_tr = high[0] - low[0];
+        self.seed_buf.push(sum_tr);
+        for i in 1..p {
+            let (h, l) = (high[i], low[i]);
+            let tr = (h - l)
+                .max((h - prev_close).abs())
+                .max((l - prev_close).abs());
+            prev_close = close[i];
+            self.seed_buf.push(tr);
+            sum_tr += tr;
+        }
+        let mut avg = sum_tr / p as f64;
+        out.push(avg);
+        // Steady state: Wilder smoothing, reciprocal hoisted out of the loop.
+        for i in p..n {
+            let (h, l) = (high[i], low[i]);
+            let tr = (h - l)
+                .max((h - prev_close).abs())
+                .max((l - prev_close).abs());
+            prev_close = close[i];
+            avg = avg.mul_add(self.n_minus_1, tr) * self.inv_period;
+            out.push(avg);
+        }
+
+        // Leave state where a full `update` replay would (seeded; seed_buf retained).
+        self.prev_close = Some(prev_close);
+        self.avg = avg;
+        self.seeded = true;
+        out
+    }
 }

 impl Indicator for Atr {
@@ -266,6 +327,81 @@ mod tests {
        }
    }

+    fn bits_eq(a: &[f64], b: &[f64]) -> bool {
+        a.len() == b.len()
+            && a.iter()
+                .zip(b)
+                .all(|(x, y)| x == y || (x.is_nan() && y.is_nan()))
+    }
+
+    fn atr_replay(period: usize, high: &[f64], low: &[f64], close: &[f64]) -> Vec<f64> {
+        let mut a = Atr::new(period).unwrap();
+        (0..high.len())
+            .map(|i| {
+                let candle = Candle::new_unchecked(close[i], high[i], low[i], close[i], 0.0, 0);
+                a.update(candle).unwrap_or(f64::NAN)
+            })
+            .collect()
+    }
+
+    /// Valid OHLC columns from a wandering base price.
+    fn columns(n: usize) -> (Vec<f64>, Vec<f64>, Vec<f64>) {
+        let base: Vec<f64> = (0..n)
+            .map(|i| (f64::from(u32::try_from(i).unwrap()) * 0.3).sin() * 5.0 + 100.0)
+            .collect();
+        let high = base.iter().map(|b| b + 1.0).collect();
+        let low = base.iter().map(|b| b - 1.0).collect();
+        (high, low, base)
+    }
+
+    #[test]
+    fn batch_atr_fast_path_is_bit_identical() {
+        let (high, low, close) = columns(300);
+        let mut atr = Atr::new(14).unwrap();
+        let got = atr.batch_atr(&high, &low, &close);
+        assert!(bits_eq(&got, &atr_replay(14, &high, &low, &close)));
+        let mut ref_atr = Atr::new(14).unwrap();
+        for i in 0..high.len() {
+            ref_atr.update(Candle::new_unchecked(
+                close[i], high[i], low[i], close[i], 0.0, 0,
+            ));
+        }
+        let next = Candle::new_unchecked(101.0, 102.0, 100.0, 101.0, 0.0, 0);
+        assert_eq!(atr.update(next), ref_atr.update(next));
+    }
+
+    #[test]
+    fn batch_atr_falls_back_when_not_fresh() {
+        let (high, low, close) = columns(40);
+        let mut atr = Atr::new(14).unwrap();
+        atr.update(Candle::new_unchecked(
+            close[0], high[0], low[0], close[0], 0.0, 0,
+        ));
+        let mut ref_atr = Atr::new(14).unwrap();
+        ref_atr.update(Candle::new_unchecked(
+            close[0], high[0], low[0], close[0], 0.0, 0,
+        ));
+        let want: Vec<f64> = (0..high.len())
+            .map(|i| {
+                ref_atr
+                    .update(Candle::new_unchecked(
+                        close[i], high[i], low[i], close[i], 0.0, 0,
+                    ))
+                    .unwrap_or(f64::NAN)
+            })
+            .collect();
+        assert!(bits_eq(&atr.batch_atr(&high, &low, &close), &want));
+    }
+
+    #[test]
+    fn batch_atr_sub_period_slice_falls_back() {
+        let (high, low, close) = columns(5);
+        let mut atr = Atr::new(14).unwrap();
+        let got = atr.batch_atr(&high, &low, &close);
+        assert!(bits_eq(&got, &atr_replay(14, &high, &low, &close)));
+        assert!(got.iter().all(|x| x.is_nan()));
+    }
+
    proptest::proptest! {
        #![proptest_config(proptest::test_runner::Config::with_cases(48))]
        #[test]
@@ -108,6 +108,82 @@ impl BollingerBands {
        self.multiplier
    }

+    /// Vectorized flat batch for bindings: returns `n * 4` values laid out as
+    /// `[upper, middle, lower, stddev]` per input row, warmup rows all `NaN`.
+    ///
+    /// For a fresh, all-finite slice it inlines `update`'s rolling `sum`/`sum_sq`
+    /// and drift-reseed, writing the four band values directly instead of an
+    /// `Option<BollingerOutput>` per element. Same add/subtract order, same reseed
+    /// cadence, same variance/`sqrt` math — so it is *bit-for-bit* equal to
+    /// replaying `update`, including the long-stream drift bound. Any other state,
+    /// or a non-finite element, defers to the exact `update` replay.
+    ///
+    /// This is a *separate* entry point from the trait [`batch`](crate::BatchExt::batch),
+    /// which returns `Vec<Option<BollingerOutput>>`; only the bindings, which want
+    /// a flat `f64` buffer, call this.
+    pub fn batch_bands(&mut self, inputs: &[f64]) -> Vec<f64> {
+        let p = self.period;
+        let n = inputs.len();
+        if self.count != 0
+            || self.updates_since_recompute != 0
+            || !inputs.iter().all(|x| x.is_finite())
+        {
+            // Slow path: exact replay of `update` into the flat layout.
+            let mut out = vec![f64::NAN; n * 4];
+            for (i, &x) in inputs.iter().enumerate() {
+                if let Some(o) = self.update(x) {
+                    out[i * 4] = o.upper;
+                    out[i * 4 + 1] = o.middle;
+                    out[i * 4 + 2] = o.lower;
+                    out[i * 4 + 3] = o.stddev;
+                }
+            }
+            return out;
+        }
+
+        let p_f64 = p as f64;
+        let mult = self.multiplier;
+        // Pre-sized output: warmup rows stay NaN, ready rows are written in place
+        // by index — no per-row `push` length/capacity check.
+        let mut out = vec![f64::NAN; n * 4];
+        for (i, &x) in inputs.iter().enumerate() {
+            if self.count == p {
+                let old = self.buf[self.head];
+                self.sum -= old;
+                self.sum_sq -= old * old;
+                self.buf[self.head] = x;
+                self.sum += x;
+                self.sum_sq += x * x;
+            } else {
+                self.buf[self.head] = x;
+                self.sum += x;
+                self.sum_sq += x * x;
+                self.count += 1;
+            }
+            self.head += 1;
+            if self.head == p {
+                self.head = 0;
+            }
+            self.updates_since_recompute += 1;
+            if self.updates_since_recompute >= RECOMPUTE_EVERY * p {
+                let chronological = self.buf[self.head..].iter().chain(&self.buf[..self.head]);
+                self.sum = chronological.clone().copied().sum();
+                self.sum_sq = chronological.map(|&v| v * v).sum();
+                self.updates_since_recompute = 0;
+            }
+            if self.count == p {
+                let mean = self.sum / p_f64;
+                let stddev = (self.sum_sq / p_f64 - mean * mean).max(0.0).sqrt();
+                let band = mult * stddev;
+                out[i * 4] = mean + band;
+                out[i * 4 + 1] = mean;
+                out[i * 4 + 2] = mean - band;
+                out[i * 4 + 3] = stddev;
+            }
+        }
+        out
+    }
+
    fn current(&self) -> Option<BollingerOutput> {
        if self.count != self.period {
            return None;
@@ -352,6 +428,79 @@ mod tests {
        );
    }

+    fn bits_eq(a: &[f64], b: &[f64]) -> bool {
+        a.len() == b.len()
+            && a.iter()
+                .zip(b)
+                .all(|(x, y)| x == y || (x.is_nan() && y.is_nan()))
+    }
+
+    /// Flat `n*4` `[upper, middle, lower, stddev]` replay of `update`.
+    fn bb_replay(period: usize, mult: f64, series: &[f64]) -> Vec<f64> {
+        let mut bb = BollingerBands::new(period, mult).unwrap();
+        let mut out = Vec::with_capacity(series.len() * 4);
+        for &x in series {
+            match bb.update(x) {
+                Some(o) => out.extend_from_slice(&[o.upper, o.middle, o.lower, o.stddev]),
+                None => out.extend_from_slice(&[f64::NAN; 4]),
+            }
+        }
+        out
+    }
+
+    #[test]
+    fn batch_bands_fast_path_is_bit_identical_with_reseed() {
+        // > 16*period inputs so the drift-reseed branch fires inside batch_bands.
+        let series: Vec<f64> = (0..500)
+            .map(|i| (f64::from(i) * 0.2).sin() * 10.0 + 50.0)
+            .collect();
+        let mut bb = BollingerBands::new(20, 2.0).unwrap();
+        let got = bb.batch_bands(&series);
+        assert!(bits_eq(&got, &bb_replay(20, 2.0, &series)));
+        // State continues identically.
+        let mut ref_bb = BollingerBands::new(20, 2.0).unwrap();
+        for &x in &series {
+            ref_bb.update(x);
+        }
+        assert_eq!(bb.update(55.0), ref_bb.update(55.0));
+    }
+
+    #[test]
+    fn batch_bands_falls_back_on_non_finite() {
+        let series = [1.0, 2.0, 3.0, f64::NAN, 5.0, 6.0, 7.0];
+        let mut bb = BollingerBands::new(3, 2.0).unwrap();
+        assert!(bits_eq(
+            &bb.batch_bands(&series),
+            &bb_replay(3, 2.0, &series)
+        ));
+    }
+
+    #[test]
+    fn batch_bands_falls_back_when_not_fresh() {
+        let mut bb = BollingerBands::new(3, 2.0).unwrap();
+        bb.update(99.0);
+        let series = [1.0, 2.0, 3.0, 4.0];
+        let mut ref_bb = BollingerBands::new(3, 2.0).unwrap();
+        ref_bb.update(99.0);
+        let mut want = Vec::new();
+        for &x in &series {
+            match ref_bb.update(x) {
+                Some(o) => want.extend_from_slice(&[o.upper, o.middle, o.lower, o.stddev]),
+                None => want.extend_from_slice(&[f64::NAN; 4]),
+            }
+        }
+        assert!(bits_eq(&bb.batch_bands(&series), &want));
+    }
+
+    #[test]
+    fn batch_bands_sub_period_slice_is_all_nan() {
+        let series = [1.0, 2.0, 3.0];
+        let mut bb = BollingerBands::new(10, 2.0).unwrap();
+        let got = bb.batch_bands(&series);
+        assert!(bits_eq(&got, &bb_replay(10, 2.0, &series)));
+        assert!(got.iter().all(|x| x.is_nan()) && got.len() == 12);
+    }
+
    #[test]
    fn ignores_non_finite_input() {
        let mut bb = BollingerBands::new(5, 2.0).unwrap();
@@ -102,6 +102,68 @@ impl Ema {
        }
    }

+    /// Whether the EMA has seen no input yet (neither seeded nor mid-warmup).
+    /// Lets composite indicators (e.g. MACD) decide if a fast batch path is safe.
+    pub(crate) fn is_fresh(&self) -> bool {
+        !self.seeded && self.warmup_buf.is_empty()
+    }
+
+    /// Force the EMA into its seeded steady state with `current` as the latest
+    /// value. Used by composite fused batch paths (MACD) to leave each sub-EMA
+    /// where a per-tick `update` replay would, so a later `update` continues
+    /// correctly. The post-seed recurrence never re-reads `warmup_buf`, so it is
+    /// left as-is.
+    pub(crate) fn seed_to(&mut self, current: f64) {
+        self.current = current;
+        self.seeded = true;
+    }
+
+    /// Vectorized batch returning one `f64` per input (`NaN` during warmup).
+    ///
+    /// Shadows the generic [`BatchNanExt::batch_nan`](crate::BatchNanExt) blanket
+    /// default via inherent-method resolution. For a fresh indicator over an
+    /// all-finite slice it runs the seed (mean of the first `period`) once and
+    /// then the bare `alpha * x + (1 - alpha) * prev` recurrence in a tight loop
+    /// with no per-element `is_finite`/`seeded` branch and no `Option` — yet uses
+    /// the identical `mul_add`, so the result is *bit-for-bit* equal to replaying
+    /// `update`. Any other state, or a non-finite element, defers to the exact
+    /// `update` replay.
+    pub fn batch_nan(&mut self, inputs: &[f64]) -> Vec<f64> {
+        let p = self.period;
+        if self.seeded || !self.warmup_buf.is_empty() || !inputs.iter().all(|x| x.is_finite()) {
+            return inputs
+                .iter()
+                .map(|&x| self.update(x).unwrap_or(f64::NAN))
+                .collect();
+        }
+
+        let n = inputs.len();
+        if n < p {
+            // Not enough to seed; mirror `update` stashing inputs for warmup.
+            self.warmup_buf.extend_from_slice(inputs);
+            return vec![f64::NAN; n];
+        }
+
+        // Warmup `[0, p-1)` is `NaN`; values from the seed on are pushed once each.
+        let mut out = vec![f64::NAN; p - 1];
+        out.reserve(n - (p - 1));
+        let seed = inputs[..p].iter().copied().sum::<f64>() / p as f64;
+        let mut cur = seed;
+        out.push(seed);
+        let (alpha, oma) = (self.alpha, self.one_minus_alpha);
+        for &x in &inputs[p..] {
+            cur = alpha.mul_add(x, oma * cur);
+            out.push(cur);
+        }
+
+        // Leave state exactly where `update` would: seeded on `current`, with the
+        // first `period` inputs retained in `warmup_buf` (never cleared post-seed).
+        self.current = cur;
+        self.seeded = true;
+        self.warmup_buf.extend_from_slice(&inputs[..p]);
+        out
+    }
+
    /// Internal helper that feeds a value without finiteness validation. The caller
    /// guarantees `input.is_finite()`. Used by MACD which has already validated.
    pub(crate) fn step_unchecked(&mut self, input: f64) -> Option<f64> {
@@ -288,6 +350,71 @@ mod tests {
        assert_eq!(ema.update(f64::INFINITY), before);
    }

+    fn bits_eq(a: &[f64], b: &[f64]) -> bool {
+        a.len() == b.len()
+            && a.iter()
+                .zip(b)
+                .all(|(x, y)| x == y || (x.is_nan() && y.is_nan()))
+    }
+
+    fn ema_replay(period: usize, series: &[f64]) -> Vec<f64> {
+        let mut e = Ema::new(period).unwrap();
+        series
+            .iter()
+            .map(|&x| e.update(x).unwrap_or(f64::NAN))
+            .collect()
+    }
+
+    #[test]
+    fn batch_nan_fast_path_is_bit_identical() {
+        let series: Vec<f64> = (0..300)
+            .map(|i| (f64::from(i) * 0.25).cos() * 8.0 + 40.0)
+            .collect();
+        let mut ema = Ema::new(14).unwrap();
+        let got = ema.batch_nan(&series);
+        assert!(bits_eq(&got, &ema_replay(14, &series)));
+        let mut ref_ema = Ema::new(14).unwrap();
+        for &x in &series {
+            ref_ema.update(x);
+        }
+        assert_eq!(ema.update(7.5), ref_ema.update(7.5));
+    }
+
+    #[test]
+    fn batch_nan_falls_back_on_non_finite() {
+        let series = [1.0, 2.0, 3.0, f64::INFINITY, 5.0, 6.0, 7.0];
+        let mut ema = Ema::new(3).unwrap();
+        assert!(bits_eq(&ema.batch_nan(&series), &ema_replay(3, &series)));
+    }
+
+    #[test]
+    fn batch_nan_falls_back_when_warming() {
+        let mut ema = Ema::new(3).unwrap();
+        ema.update(10.0); // mid-warmup: warmup_buf non-empty, not seeded
+        let series = [1.0, 2.0, 3.0, 4.0];
+        let mut ref_ema = Ema::new(3).unwrap();
+        ref_ema.update(10.0);
+        let want: Vec<f64> = series
+            .iter()
+            .map(|&x| ref_ema.update(x).unwrap_or(f64::NAN))
+            .collect();
+        assert!(bits_eq(&ema.batch_nan(&series), &want));
+    }
+
+    #[test]
+    fn batch_nan_sub_period_slice_stays_unseeded() {
+        let series = [1.0, 2.0];
+        let mut ema = Ema::new(5).unwrap();
+        let got = ema.batch_nan(&series);
+        assert!(got.iter().all(|x| x.is_nan()) && got.len() == 2);
+        assert!(!ema.is_ready());
+        // Warmup state was stashed: feeding the rest seeds exactly as a full stream.
+        assert!(bits_eq(
+            &[ema.update(3.0).unwrap_or(f64::NAN)],
+            &[ema_replay(5, &[1.0, 2.0, 3.0])[2]]
+        ));
+    }
+
    proptest::proptest! {
        #![proptest_config(proptest::test_runner::Config::with_cases(48))]
        #[test]
@@ -86,6 +86,116 @@ impl MacdIndicator {
    pub const fn value(&self) -> Option<MacdOutput> {
        self.last
    }
+
+    /// Vectorized flat batch for bindings: `n * 3` values laid out as
+    /// `[macd, signal, histogram]` per input row, warmup rows all `NaN`.
+    ///
+    /// For a fresh, all-finite slice long enough for a full output it runs the
+    /// fast EMA, slow EMA and signal EMA as three recurrences fused into a single
+    /// pass with one allocation — no `Option` per tick, no per-EMA intermediate
+    /// buffers, identical SMA-mean seeds (division) and `mul_add` recurrences. The
+    /// result is *bit-for-bit* equal to replaying `update`. Anything else (not
+    /// fresh, non-finite, or too short to emit) defers to the exact `update`
+    /// replay.
+    ///
+    /// Separate from the trait [`batch`](crate::BatchExt::batch), which stays a
+    /// bit-identical `update` replay; only the bindings call this.
+    pub fn batch_macd(&mut self, inputs: &[f64]) -> Vec<f64> {
+        let n = inputs.len();
+        let (fp, sp, gp) = (self.fast_period, self.slow_period, self.signal_period);
+        // First full output needs the slow EMA seeded (index sp-1) plus gp signal
+        // values: index sp + gp - 2. Below that, or non-fresh/non-finite, replay.
+        if self.last.is_some()
+            || !self.fast.is_fresh()
+            || !self.slow.is_fresh()
+            || !self.signal_ema.is_fresh()
+            || n < sp + gp - 1
+            || !inputs.iter().all(|x| x.is_finite())
+        {
+            let mut out = vec![f64::NAN; n * 3];
+            for (i, &x) in inputs.iter().enumerate() {
+                if let Some(o) = self.update(x) {
+                    out[i * 3] = o.macd;
+                    out[i * 3 + 1] = o.signal;
+                    out[i * 3 + 2] = o.histogram;
+                }
+            }
+            return out;
+        }
+
+        // Pre-sized output: warmup rows stay NaN, full-output rows are written in
+        // place by index — no per-row `push` length/capacity check.
+        let mut out = vec![f64::NAN; n * 3];
+        let (fa, fo) = (self.fast.alpha(), 1.0 - self.fast.alpha());
+        let (sa, so) = (self.slow.alpha(), 1.0 - self.slow.alpha());
+        let (ga, go) = (self.signal_ema.alpha(), 1.0 - self.signal_ema.alpha());
+        let (fp_f, sp_f, gp_f) = (fp as f64, sp as f64, gp as f64);
+
+        let (mut fast_val, mut slow_val, mut sig) = (0.0_f64, 0.0_f64, 0.0_f64);
+        let (mut fsum, mut ssum, mut gsum) = (0.0_f64, 0.0_f64, 0.0_f64);
+        let mut sig_count = 0usize; // signal-EMA seed progress (raw MACD values seen)
+        let mut sig_seeded = false;
+        let mut last = MacdOutput {
+            macd: 0.0,
+            signal: 0.0,
+            histogram: 0.0,
+        };
+
+        for (i, &x) in inputs.iter().enumerate() {
+            // Fast EMA: SMA-seeded at index fp-1, then recurrence.
+            if i < fp {
+                fsum += x;
+                if i == fp - 1 {
+                    fast_val = fsum / fp_f;
+                }
+            } else {
+                fast_val = fa.mul_add(x, fo * fast_val);
+            }
+            // Slow EMA: SMA-seeded at index sp-1, then recurrence.
+            if i < sp {
+                ssum += x;
+                if i == sp - 1 {
+                    slow_val = ssum / sp_f;
+                }
+            } else {
+                slow_val = sa.mul_add(x, so * slow_val);
+            }
+            if i + 1 < sp {
+                continue; // slow EMA not seeded yet → no raw MACD line
+            }
+            let macd = fast_val - slow_val;
+            // Signal EMA over the MACD line: SMA-seeded over its first gp values.
+            let signal = if sig_seeded {
+                sig = ga.mul_add(macd, go * sig);
+                sig
+            } else {
+                gsum += macd;
+                sig_count += 1;
+                if sig_count < gp {
+                    continue; // signal EMA still seeding → no full output
+                }
+                sig = gsum / gp_f;
+                sig_seeded = true;
+                sig
+            };
+            let histogram = macd - signal;
+            out[i * 3] = macd;
+            out[i * 3 + 1] = signal;
+            out[i * 3 + 2] = histogram;
+            last = MacdOutput {
+                macd,
+                signal,
+                histogram,
+            };
+        }
+
+        // Leave every sub-EMA and `last` where a full `update` replay would.
+        self.fast.seed_to(fast_val);
+        self.slow.seed_to(slow_val);
+        self.signal_ema.seed_to(sig);
+        self.last = Some(last);
+        out
+    }
 }

 impl Indicator for MacdIndicator {
@@ -256,6 +366,79 @@ mod tests {
        assert_eq!(macd.update(1.0), None);
    }

+    fn bits_eq(a: &[f64], b: &[f64]) -> bool {
+        a.len() == b.len()
+            && a.iter()
+                .zip(b)
+                .all(|(x, y)| x == y || (x.is_nan() && y.is_nan()))
+    }
+
+    /// Flat `n*3` `[macd, signal, histogram]` replay of `update`.
+    fn macd_replay(series: &[f64]) -> Vec<f64> {
+        let mut m = MacdIndicator::classic();
+        let mut out = Vec::with_capacity(series.len() * 3);
+        for &x in series {
+            match m.update(x) {
+                Some(o) => out.extend_from_slice(&[o.macd, o.signal, o.histogram]),
+                None => out.extend_from_slice(&[f64::NAN; 3]),
+            }
+        }
+        out
+    }
+
+    #[test]
+    fn batch_macd_fast_path_is_bit_identical() {
+        let series: Vec<f64> = (0..300)
+            .map(|i| (f64::from(i) * 0.4).cos() * 10.0 + 100.0)
+            .collect();
+        let mut macd = MacdIndicator::classic();
+        let got = macd.batch_macd(&series);
+        assert!(bits_eq(&got, &macd_replay(&series)));
+        // Sub-EMA + last state left where the replay would: continued update agrees.
+        let mut ref_macd = MacdIndicator::classic();
+        for &x in &series {
+            ref_macd.update(x);
+        }
+        let (a, b) = (macd.update(101.0), ref_macd.update(101.0));
+        assert_eq!(a.is_some(), b.is_some());
+        assert_relative_eq!(a.unwrap().macd, b.unwrap().macd, epsilon = 1e-12);
+    }
+
+    #[test]
+    fn batch_macd_falls_back_on_non_finite() {
+        let mut series: Vec<f64> = (0..60).map(|i| f64::from(i) + 100.0).collect();
+        series[40] = f64::NAN;
+        let mut macd = MacdIndicator::classic();
+        assert!(bits_eq(&macd.batch_macd(&series), &macd_replay(&series)));
+    }
+
+    #[test]
+    fn batch_macd_falls_back_when_not_fresh() {
+        let series: Vec<f64> = (0..60).map(|i| f64::from(i) + 100.0).collect();
+        let mut macd = MacdIndicator::classic();
+        macd.update(50.0);
+        let mut ref_macd = MacdIndicator::classic();
+        ref_macd.update(50.0);
+        let mut want = Vec::new();
+        for &x in &series {
+            match ref_macd.update(x) {
+                Some(o) => want.extend_from_slice(&[o.macd, o.signal, o.histogram]),
+                None => want.extend_from_slice(&[f64::NAN; 3]),
+            }
+        }
+        assert!(bits_eq(&macd.batch_macd(&series), &want));
+    }
+
+    #[test]
+    fn batch_macd_too_short_for_output_falls_back() {
+        // n < slow + signal - 1 (= 34): no full output, routed to the replay.
+        let series: Vec<f64> = (0..20).map(|i| f64::from(i) + 100.0).collect();
+        let mut macd = MacdIndicator::classic();
+        let got = macd.batch_macd(&series);
+        assert!(bits_eq(&got, &macd_replay(&series)));
+        assert!(got.iter().all(|x| x.is_nan()));
+    }
+
    #[test]
    fn ignores_non_finite_input() {
        let mut macd = MacdIndicator::classic();
@@ -81,6 +81,76 @@ impl Rsi {
        self.last_value
    }

+    /// Vectorized batch returning one `f64` per input (`NaN` during warmup).
+    ///
+    /// Shadows the generic [`BatchNanExt::batch_nan`](crate::BatchNanExt) blanket
+    /// default. RSI is a recursive (IIR) filter — Wilder smoothing — so it cannot
+    /// be SIMD-vectorized any more than the C peers manage; the win is purely in
+    /// stripping per-tick overhead. For a fresh indicator over an all-finite slice
+    /// long enough to seed (`n > period`) it runs the seed once and then the bare
+    /// smoothing recurrence in a tight loop with no per-tick `is_finite`/`has_prev`/
+    /// `avgs_seeded` branch and no `Option`, using the identical division at the
+    /// seed and `mul_add`/`rsi_from_avgs` afterwards — so it is *bit-for-bit* equal
+    /// to replaying `update`. Shorter or non-fresh/non-finite inputs defer to the
+    /// exact `update` replay.
+    pub fn batch_nan(&mut self, inputs: &[f64]) -> Vec<f64> {
+        let p = self.period;
+        let n = inputs.len();
+        if self.has_prev
+            || self.avgs_seeded
+            || !self.seed_buf_gains.is_empty()
+            || n <= p
+            || !inputs.iter().all(|x| x.is_finite())
+        {
+            return inputs
+                .iter()
+                .map(|&x| self.update(x).unwrap_or(f64::NAN))
+                .collect();
+        }
+
+        // Warmup `[0, p)` is `NaN`; outputs from index `p` on are pushed once each.
+        let mut out = vec![f64::NAN; p];
+        out.reserve(n - p);
+        // Seed from the first `period` diffs (inputs[1..=p]); index 0 only sets the
+        // baseline. Retain the seed gains/losses exactly as `update` leaves them.
+        let mut prev = inputs[0];
+        let (mut sum_gain, mut sum_loss) = (0.0_f64, 0.0_f64);
+        for &x in &inputs[1..=p] {
+            let diff = x - prev;
+            prev = x;
+            let gain = if diff > 0.0 { diff } else { 0.0 };
+            let loss = if diff < 0.0 { -diff } else { 0.0 };
+            self.seed_buf_gains.push(gain);
+            self.seed_buf_losses.push(loss);
+            sum_gain += gain;
+            sum_loss += loss;
+        }
+        let p_f64 = p as f64;
+        let mut ag = sum_gain / p_f64;
+        let mut al = sum_loss / p_f64;
+        out.push(Self::rsi_from_avgs(ag, al));
+
+        // Steady state: Wilder smoothing, reciprocal hoisted, one `rsi_from_avgs`.
+        for &x in &inputs[p + 1..] {
+            let diff = x - prev;
+            prev = x;
+            let gain = if diff > 0.0 { diff } else { 0.0 };
+            let loss = if diff < 0.0 { -diff } else { 0.0 };
+            ag = ag.mul_add(self.n_minus_1, gain) * self.inv_period;
+            al = al.mul_add(self.n_minus_1, loss) * self.inv_period;
+            out.push(Self::rsi_from_avgs(ag, al));
+        }
+
+        // Leave state where a full `update` replay would.
+        self.prev_close = prev;
+        self.has_prev = true;
+        self.avg_gain = ag;
+        self.avg_loss = al;
+        self.avgs_seeded = true;
+        self.last_value = Some(out[n - 1]);
+        out
+    }
+
    fn rsi_from_avgs(avg_gain: f64, avg_loss: f64) -> f64 {
        // Algebraically `100 - 100/(1 + ag/al)` collapses to `100·ag/(ag+al)`,
        // which needs a single division instead of two and removes the separate
@@ -376,6 +446,65 @@ mod tests {
        assert_eq!(rsi.value(), before);
    }

+    fn bits_eq(a: &[f64], b: &[f64]) -> bool {
+        a.len() == b.len()
+            && a.iter()
+                .zip(b)
+                .all(|(x, y)| x == y || (x.is_nan() && y.is_nan()))
+    }
+
+    fn rsi_replay(period: usize, series: &[f64]) -> Vec<f64> {
+        let mut r = Rsi::new(period).unwrap();
+        series
+            .iter()
+            .map(|&x| r.update(x).unwrap_or(f64::NAN))
+            .collect()
+    }
+
+    #[test]
+    fn batch_nan_fast_path_is_bit_identical() {
+        let series: Vec<f64> = (0..300)
+            .map(|i| (f64::from(i) * 0.3).sin() * 5.0 + f64::from(i) * 0.1 + 100.0)
+            .collect();
+        let mut rsi = Rsi::new(14).unwrap();
+        let got = rsi.batch_nan(&series);
+        assert!(bits_eq(&got, &rsi_replay(14, &series)));
+        let mut ref_rsi = Rsi::new(14).unwrap();
+        for &x in &series {
+            ref_rsi.update(x);
+        }
+        assert_eq!(rsi.update(123.0), ref_rsi.update(123.0));
+    }
+
+    #[test]
+    fn batch_nan_falls_back_on_non_finite() {
+        let series = [10.0, 11.0, 9.0, f64::NAN, 12.0, 13.0, 8.0];
+        let mut rsi = Rsi::new(3).unwrap();
+        assert!(bits_eq(&rsi.batch_nan(&series), &rsi_replay(3, &series)));
+    }
+
+    #[test]
+    fn batch_nan_falls_back_when_not_fresh() {
+        let mut rsi = Rsi::new(3).unwrap();
+        rsi.update(50.0);
+        let series = [51.0, 49.0, 52.0, 53.0, 50.0];
+        let mut ref_rsi = Rsi::new(3).unwrap();
+        ref_rsi.update(50.0);
+        let want: Vec<f64> = series
+            .iter()
+            .map(|&x| ref_rsi.update(x).unwrap_or(f64::NAN))
+            .collect();
+        assert!(bits_eq(&rsi.batch_nan(&series), &want));
+    }
+
+    #[test]
+    fn batch_nan_too_short_to_seed_falls_back() {
+        // n <= period: routed to the exact replay (cannot seed yet).
+        let series = [10.0, 11.0, 12.0];
+        let mut rsi = Rsi::new(3).unwrap();
+        assert!(bits_eq(&rsi.batch_nan(&series), &rsi_replay(3, &series)));
+    }
+
    proptest::proptest! {
        #![proptest_config(proptest::test_runner::Config::with_cases(48))]
        #[test]
@@ -86,6 +86,62 @@ impl Sma {
            None
        }
    }
+
+    /// Vectorized batch returning one `f64` per input (`NaN` during warmup).
+    ///
+    /// Shadows the generic [`BatchNanExt::batch_nan`](crate::BatchNanExt) blanket
+    /// default via inherent-method resolution. For a fresh, all-finite slice it
+    /// inlines `update`'s rolling sum and drift-reseed, writing the mean as a bare
+    /// `f64` (warmup → `NaN`) instead of allocating an `Option<f64>` per element
+    /// and walking the result a second time. Same add/subtract order, same reseed
+    /// cadence, same `sum / period` division — so it is *bit-for-bit* equal to
+    /// replaying `update`, including the long-stream drift bound. Any other state,
+    /// or a non-finite element, defers to the exact `update` replay.
+    pub fn batch_nan(&mut self, inputs: &[f64]) -> Vec<f64> {
+        let p = self.period;
+        if self.count != 0
+            || self.updates_since_recompute != 0
+            || !inputs.iter().all(|x| x.is_finite())
+        {
+            return inputs
+                .iter()
+                .map(|&x| self.update(x).unwrap_or(f64::NAN))
+                .collect();
+        }
+
+        let p_f64 = p as f64;
+        let mut out = Vec::with_capacity(inputs.len());
+        for &x in inputs {
+            if self.count == p {
+                self.sum -= self.buf[self.head];
+                self.buf[self.head] = x;
+                self.sum += x;
+            } else {
+                self.buf[self.head] = x;
+                self.sum += x;
+                self.count += 1;
+            }
+            self.head += 1;
+            if self.head == p {
+                self.head = 0;
+            }
+            self.updates_since_recompute += 1;
+            if self.updates_since_recompute >= RECOMPUTE_EVERY * p {
+                self.sum = self.buf[self.head..]
+                    .iter()
+                    .chain(&self.buf[..self.head])
+                    .copied()
+                    .sum();
+                self.updates_since_recompute = 0;
+            }
+            out.push(if self.count == p {
+                self.sum / p_f64
+            } else {
+                f64::NAN
+            });
+        }
+        out
+    }
 }

 impl Indicator for Sma {
@@ -246,6 +302,69 @@ mod tests {
        }
    }

+    /// NaN-aware bit-equality for the `f64`-with-NaN-warmup batch outputs.
+    fn bits_eq(a: &[f64], b: &[f64]) -> bool {
+        a.len() == b.len()
+            && a.iter()
+                .zip(b)
+                .all(|(x, y)| x == y || (x.is_nan() && y.is_nan()))
+    }
+
+    fn sma_replay(period: usize, series: &[f64]) -> Vec<f64> {
+        let mut s = Sma::new(period).unwrap();
+        series
+            .iter()
+            .map(|&x| s.update(x).unwrap_or(f64::NAN))
+            .collect()
+    }
+
+    #[test]
+    fn batch_nan_fast_path_is_bit_identical_with_reseed() {
+        // > 16*period inputs so the drift-reseed branch fires inside batch_nan.
+        let series: Vec<f64> = (0..500)
+            .map(|i| (f64::from(i) * 0.2).sin() * 10.0 + 50.0)
+            .collect();
+        let mut sma = Sma::new(14).unwrap();
+        let got = sma.batch_nan(&series);
+        assert!(bits_eq(&got, &sma_replay(14, &series)));
+        // State left where the replay would: continued updates agree.
+        let mut ref_sma = Sma::new(14).unwrap();
+        for &x in &series {
+            ref_sma.update(x);
+        }
+        assert_eq!(sma.update(42.0), ref_sma.update(42.0));
+    }
+
+    #[test]
+    fn batch_nan_falls_back_on_non_finite() {
+        let series = [1.0, 2.0, f64::NAN, 4.0, 5.0, 6.0];
+        let mut sma = Sma::new(3).unwrap();
+        assert!(bits_eq(&sma.batch_nan(&series), &sma_replay(3, &series)));
+    }
+
+    #[test]
+    fn batch_nan_falls_back_when_not_fresh() {
+        let mut sma = Sma::new(3).unwrap();
+        sma.update(99.0);
+        let series = [1.0, 2.0, 3.0, 4.0];
+        let mut ref_sma = Sma::new(3).unwrap();
+        ref_sma.update(99.0);
+        let want: Vec<f64> = series
+            .iter()
+            .map(|&x| ref_sma.update(x).unwrap_or(f64::NAN))
+            .collect();
+        assert!(bits_eq(&sma.batch_nan(&series), &want));
+    }
+
+    #[test]
+    fn batch_nan_sub_period_slice_is_all_nan() {
+        let series = [1.0, 2.0, 3.0];
+        let mut sma = Sma::new(10).unwrap();
+        let got = sma.batch_nan(&series);
+        assert!(bits_eq(&got, &sma_replay(10, &series)));
+        assert!(got.iter().all(|x| x.is_nan()));
+    }
+
    proptest::proptest! {
        #![proptest_config(proptest::test_runner::Config::with_cases(64))]
        #[test]
@@ -166,4 +166,4 @@ pub use indicators::PnfColumn;
 pub use indicators::RenkoBrick;
 pub use microstructure::{Level, OrderBook, Side, Trade, TradeQuote};
 pub use ohlcv::{Candle, Tick};
-pub use traits::{BarBuilder, BatchExt, Chain, Indicator};
+pub use traits::{BarBuilder, BatchExt, BatchNanExt, Chain, Indicator};
@@ -90,6 +90,29 @@ pub trait BatchExt: Indicator {

 impl<T: Indicator> BatchExt for T {}

+/// Fast batch for scalar `f64 -> f64` indicators.
+///
+/// The generic [`BatchExt::batch`] returns `Vec<Option<f64>>` — 16 bytes per
+/// element (no niche fits an arbitrary `f64`), which a caller wanting a dense
+/// `f64` series then has to walk a second time to map warmup `None`s to `NaN`.
+/// This skips both the wide intermediate and the second pass: one allocation,
+/// one pass, warmup encoded as `NaN`. The default body is bit-identical to
+/// replaying `update`; indicators with a vectorizable closed form override it
+/// with an inherent `batch_nan` of the same name, which wins method resolution
+/// over this trait default.
+pub trait BatchNanExt: Indicator<Input = f64, Output = f64> {
+    /// One `f64` per input, warmup positions filled with `NaN`.
+    fn batch_nan(&mut self, inputs: &[f64]) -> Vec<f64> {
+        let mut out = Vec::with_capacity(inputs.len());
+        for &x in inputs {
+            out.push(self.update(x).unwrap_or(f64::NAN));
+        }
+        out
+    }
+}
+
+impl<T: Indicator<Input = f64, Output = f64>> BatchNanExt for T {}
+
 /// A streaming *bar builder* — an alternative-chart constructor (Renko, Kagi,
 /// Point-and-Figure) that turns a candle stream into a stream of price-driven
 /// bars.
@@ -297,6 +320,17 @@ mod tests {
        assert_eq!(out, vec![Some(1.0), Some(2.0), Some(3.0)]);
    }

+    /// The blanket [`BatchNanExt::batch_nan`] default (used by every scalar
+    /// indicator without an inherent fast path) maps `update` outputs to a dense
+    /// `f64` series, warmup `None` becoming `NaN`. `Identity` is always ready, so
+    /// the result is just the inputs back.
+    #[test]
+    fn batch_nan_default_maps_none_to_nan() {
+        let mut id = Identity::default();
+        let out = id.batch_nan(&[1.0, 2.0, 3.0]);
+        assert_eq!(out, vec![1.0, 2.0, 3.0]);
+    }
+
    #[test]
    fn chain_pipes_first_into_second() {
        let mut c = Chain::new(Doubler::default(), Doubler::default());