gavindiaz/ferro-ta
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: refresh benchmark coverage and harden CI tooling

Browse Source 
Refresh the benchmark and performance surface across the repo. This updates the benchmark wrappers and helper scripts, regenerates the checked-in benchmark and perf-contract artifacts, and folds in the related roadmap, compatibility, and example notebook changes that belong with this performance-focused pass.

Harden the Python CI and local pre-push flow so the same checks pass reliably in both places. The workflow and pre-push script now use module-safe uv typecheck invocations, the Python test environment installs the optional MCP dependency needed by the MCP server tests, and one-off root benchmark outputs are ignored to keep the repo clean.

Align local tooling with the current project configuration by updating the Ruff pre-commit hook, tightening the API typing and MCP server helpers, and refreshing the lockfile to pick up the audited PyJWT fix while preserving the rest of the staged source changes.
This commit is contained in:
Pratik Bhadane
2026-03-24 14:52:20 +05:30
parent 53566b9d82
commit 71b6343e92
48 changed files with 3107 additions and 988 deletions
Download Patch File Download Diff File Expand all files Collapse all files
PERFORMANCE_ROADMAP.md
+5 -5
View File
@@ -70,23 +70,23 @@ rustflags = ["-C", "target-cpu=native"]
### Phase 3: Algorithm-Level Optimizations (Target: 5-10x improvement)
#### SMA — O(n) running sum
Current: recomputes each window.
Current: recomputes each window.
Target: single-pass running sum (already done in Rust — verify SIMD path is hit).
#### BBANDS — Welford's algorithm
Current: compute mean, then variance in two passes.
Current: compute mean, then variance in two passes.
Target: Welford's online algorithm — single pass, better cache utilization.
#### ATR/ADX — Avoid redundant True Range calculations
Current: ATR → ADX each compute TR independently.
Current: ATR → ADX each compute TR independently.
Target: Compute TR once, share with ATR, NATR, +DI, -DI, ADX in a single pass.
#### MACD — Reuse EMA computations
Current: Compute fast EMA and slow EMA separately.
Current: Compute fast EMA and slow EMA separately.
Target: Single function computes both EMAs in one pass.
#### Candlestick Patterns — Batch lookup table
Current: Sequential condition checks per bar.
Current: Sequential condition checks per bar.
Target: Pre-compute body/shadow ratios, vectorized pattern matching.
### Phase 4: Streaming Precomputation (Target: 100x for incremental updates)