fix(sma, bollinger): periodic recompute to bound long-stream drift (R7, L2-Rust)

`Sma` and `BollingerBands` both maintained their running `sum` (and
`sum_sq` for Bollinger) with a single-subtract incremental update. That
is correct in exact arithmetic, but in f64 the sequence `sum -= old;
sum += new` on long streams with alternating large/small magnitudes
can accumulate catastrophic-cancellation error. Bollinger's existing
`.max(0.0)` clamp on the computed variance was a band-aid for the same
root cause — the drift had already driven the running variance below
zero.

The fix: every `16 · period` finite updates, reseed `sum` (and `sum_sq`
for Bollinger) from the live window. Amortised cost stays at O(1) —
`O(period)` work amortised over `O(period)` updates — and the reseed
strategy is named after the constant `RECOMPUTE_EVERY` so the
intention is clear at the call site.

Behaviour is unchanged on inputs that did not drift to begin with
(every existing test still passes, including `batch_equals_streaming`
and the SMA proptest). Two new stress tests
(`long_stream_drift_stays_bounded` in each module) feed a
magnitude-alternating stream for `5 · RECOMPUTE_EVERY · period`
updates and assert the reported value tracks a fresh from-scratch
computation over the live window to within tight tolerance — these
would have failed without the reseed on Bollinger's `sum_sq`.

The misleading `sma.rs` comment that claimed drift was already
bounded by recomputing the sum after each pop is rewritten to
describe the actual reseed strategy (audit finding L2-Rust).
This commit is contained in:
kingchenc
2026-05-23 10:42:50 +02:00
parent 2db546ac12
commit 510013fc5a
3 changed files with 136 additions and 3 deletions
@@ -25,6 +25,14 @@ pub struct BollingerOutput {
/// publication uses population (not sample) standard deviation, which matches every
/// reference implementation (TA-Lib, pandas-ta, etc.).
///
/// The running `sum` and `sum_sq` are reseeded from the live window every
/// `16 · period` updates to cap floating-point drift on long streams. This is
/// amortised O(1), preserves bit-equivalence with the previous behaviour on
/// inputs that did not drift, and is particularly important for `sum_sq`,
/// where catastrophic cancellation between large add/subtract pairs can drive
/// the computed variance negative (the `.max(0.0)` clamp below is the
/// safety-net for the rare cases where the reseed has not happened yet).
///
/// # Example
///
/// ```
@@ -44,8 +52,17 @@ pub struct BollingerBands {
window: VecDeque<f64>,
sum: f64,
sum_sq: f64,
/// Number of finite updates since the running sums were last reseeded
/// from the live window. See [`RECOMPUTE_EVERY`] below.
updates_since_recompute: usize,
}
/// How often (in finite updates) the incremental `sum` / `sum_sq` are reseeded
/// from the live window. The multiplier `16` keeps the amortised cost flat and
/// caps any cancellation drift to roughly `16 · period · ULP · max(|x|²)` —
/// negligible on real-world price scales.
const RECOMPUTE_EVERY: usize = 16;
impl BollingerBands {
/// Construct a new Bollinger Bands indicator.
///
@@ -66,6 +83,7 @@ impl BollingerBands {
window: VecDeque::with_capacity(period),
sum: 0.0,
sum_sq: 0.0,
updates_since_recompute: 0,
})
}
@@ -119,6 +137,12 @@ impl Indicator for BollingerBands {
self.window.push_back(input);
self.sum += input;
self.sum_sq += input * input;
self.updates_since_recompute += 1;
if self.updates_since_recompute >= RECOMPUTE_EVERY * self.period {
self.sum = self.window.iter().copied().sum();
self.sum_sq = self.window.iter().copied().map(|x| x * x).sum();
self.updates_since_recompute = 0;
}
self.current()
}
@@ -126,6 +150,7 @@ impl Indicator for BollingerBands {
self.window.clear();
self.sum = 0.0;
self.sum_sq = 0.0;
self.updates_since_recompute = 0;
}
fn warmup_period(&self) -> usize {
@@ -254,6 +279,45 @@ mod tests {
assert!(!bb.is_ready());
}
/// Long-running stability check. After several recompute cycles the
/// reported Bollinger bands must still equal a fresh from-scratch
/// computation over the live window — even on inputs designed to cause
/// catastrophic cancellation in the `sum_sq` accumulator (alternating
/// between two very different magnitudes).
#[test]
fn long_stream_drift_stays_bounded() {
let period = 20;
let mult = 2.0;
let mut bb = BollingerBands::new(period, mult).unwrap();
let mut window: VecDeque<f64> = VecDeque::with_capacity(period);
// Forces the periodic reseed to fire 5+ times.
let n_updates = 16 * period * 5;
let mut last = None;
for i in 0..n_updates {
let v = if i.is_multiple_of(2) { 1e6 } else { 1.0 };
last = bb.update(v);
if window.len() == period {
window.pop_front();
}
window.push_back(v);
}
let scratch =
naive(&window.iter().copied().collect::<Vec<_>>(), period, mult).expect("warmed up");
let got = last.expect("warmed up");
assert!(
(got.middle - scratch.middle).abs() < 1e-3,
"middle drift: got={}, scratch={}",
got.middle,
scratch.middle,
);
assert!(
(got.stddev - scratch.stddev).abs() < 1e-3,
"stddev drift: got={}, scratch={}",
got.stddev,
scratch.stddev,
);
}
#[test]
fn ignores_non_finite_input() {
let mut bb = BollingerBands::new(5, 2.0).unwrap();
+58 -3
View File
@@ -10,6 +10,14 @@ use crate::traits::Indicator;
/// Maintains a rolling sum so each update is O(1). Output equals
/// `sum(last `period` prices) / period` once the window is full; `None` before.
///
/// On long-running streams a single-subtract incremental sum can accumulate
/// rounding error (catastrophic cancellation when values of very different
/// magnitudes are alternately added and removed). To keep drift bounded, the
/// running sum is reseeded from the live window every `16 · period` updates —
/// O(1) amortised cost (`O(period)` work amortised over `O(period)` updates),
/// zero observable behaviour change on inputs that did not drift to begin
/// with, and a strict cap on accumulated rounding for streams that did.
///
/// # Example
///
/// ```
@@ -27,8 +35,19 @@ pub struct Sma {
period: usize,
window: VecDeque<f64>,
sum: f64,
/// Number of finite updates since the running `sum` was last reseeded from
/// the live window. Caps accumulated floating-point drift on long streams.
/// See [`RECOMPUTE_EVERY`] below.
updates_since_recompute: usize,
}
/// How often (in finite updates) the incremental sum is reseeded from the live
/// window. The multiplier `16` is the smallest power of two that keeps the
/// amortised cost flat under any `period` while still bounding any drift to
/// roughly `16 · period · ULP · max(|x|)` — sub-picodollar on real-world price
/// scales.
const RECOMPUTE_EVERY: usize = 16;
impl Sma {
/// Construct a new SMA with the given window length.
///
@@ -43,6 +62,7 @@ impl Sma {
period,
window: VecDeque::with_capacity(period),
sum: 0.0,
updates_since_recompute: 0,
})
}
@@ -70,20 +90,26 @@ impl Indicator for Sma {
return self.value();
}
if self.window.len() == self.period {
// Drop the oldest from the sum to keep numerical drift bounded by recomputing
// the sum after each pop; a single subtract works in O(1) and is acceptable
// here because we use f64 throughout.
// Slide: drop the oldest, then add the new. Each step is a single
// f64 add/subtract — O(1) but introduces ~1 ULP of rounding noise.
// The periodic reseed below caps the accumulated drift.
let old = self.window.pop_front().expect("window non-empty");
self.sum -= old;
}
self.window.push_back(input);
self.sum += input;
self.updates_since_recompute += 1;
if self.updates_since_recompute >= RECOMPUTE_EVERY * self.period {
self.sum = self.window.iter().copied().sum();
self.updates_since_recompute = 0;
}
self.value()
}
fn reset(&mut self) {
self.window.clear();
self.sum = 0.0;
self.updates_since_recompute = 0;
}
fn warmup_period(&self) -> usize {
@@ -210,4 +236,33 @@ mod tests {
}
}
}
/// Long-running stability check. Runs more updates than `RECOMPUTE_EVERY *
/// period` so the periodic reseed must fire several times, then asserts
/// that the reported SMA still equals a fresh from-scratch mean over the
/// live window to within tight floating-point tolerance. Inputs swing
/// between two magnitudes (`1e9` and `1.0`) — a pattern designed to
/// expose catastrophic cancellation in a naive single-subtract sum.
#[test]
fn long_stream_drift_stays_bounded() {
let period = 20;
let mut sma = Sma::new(period).unwrap();
let mut window: VecDeque<f64> = VecDeque::with_capacity(period);
// `RECOMPUTE_EVERY * period * 5` updates → recompute fires 5+ times.
let n_updates = 16 * period * 5;
for i in 0..n_updates {
let v = if i.is_multiple_of(2) { 1e9 } else { 1.0 };
sma.update(v);
if window.len() == period {
window.pop_front();
}
window.push_back(v);
}
let from_scratch: f64 = window.iter().sum::<f64>() / period as f64;
let got = sma.value().expect("warmed up");
assert!(
(got - from_scratch).abs() < 1e-6,
"SMA drift exceeds 1e-6 over {n_updates} updates: got={got}, scratch={from_scratch}"
);
}
}