Changelog

All notable changes to this project will be documented in this file.

[Unreleased]

Fixed

Benchmark synthetic data is now reproducible across numpy versions; published benchmark numbers are reset (quantwave-5yjg).

benchmarks/data.py drew from np.random.default_rng, whose Generator streams NumPy explicitly reserves the right to change between feature releases (NEP 19 freezes only the legacy RandomState). The same seed therefore produced different data on different numpy builds — confirmed here as numpy 1.26.4 and 2.4.6 yielding different digests from an identical config. Since dataset.frame_hash in benchmarks/results/latest.json exists precisely to prove two runs measured the same data, it could not do its job, and the nightly committed a spurious diff to main whenever CI's pip resolution shifted.

Values now come from SplitMix64 implemented in explicit uint64 arithmetic inside benchmarks/data.py, depending on no library RNG. It is counter-based, so row i is a pure function of (seed, column, i) — a 100k smoke run is a true prefix of the 1M nightly run. frame_hash also folds in column name, dtype and length and normalises to little-endian, so a reordered or retyped frame can no longer collide with the original.

This changes the synthetic dataset, so benchmark figures produced before this release are not comparable with those after it. tests/python/test_benchmark_harness.py now pins the generator with golden digests; if they ever fail, the data stream moved and the baseline must be reset deliberately rather than re-blessed.

Changed

BREAKING (Polars plugin surface): ta_beta now defaults to timeperiod=5 and ta_correl to timeperiod=30, matching TA-Lib and their non-prefixed siblings (quantwave-0h4o).

The ta_* signatures were generated with a single blanket timeperiod=14. TA-Lib does not use one uniform default — BETA is 5 and CORREL is 30, while the LINEARREG family, TSF, ATR and NATR are 14 — so .ta.ta_correl(other) silently computed a 14-bar correlation where .ta.correl(other) and TA-Lib compute a 30-bar one, and .ta.ta_beta(other) a 14-bar beta where both others use 5. The formulas were always correct and the twins are bit-identical at equal periods; only the default diverged. That put the divergence on exactly the wrong surface: the ta_ prefix exists to promise TA-Lib fidelity.
```
# before — 14 bars, disagreeing with .ta.correl on the same data
pl.col("a").ta.ta_correl("b")
# now — 30 bars, TA-Lib's own default; agrees with .ta.correl("b")
pl.col("a").ta.ta_correl("b")
```
This changes results for anyone who relied on the old bare-call default. Pass timeperiod=14 explicitly to keep the previous numbers. The classic-array shim inherits its defaults from these signatures, so quantwave.talib.TABETA / TACORREL change with them and now agree with BETA / CORREL.

Audited the full ta_* set: ta_atr, ta_natr, ta_linearreg, ta_linearreg_angle, ta_linearreg_intercept, ta_linearreg_slope and ta_tsf were already at TA-Lib's 14 and are unchanged; ta_trange takes no period. Only ta_beta and ta_correl moved. The non-prefixed max, min, sum, maxindex, minindex and wma keep quantwave's house default of 14 even though TA-Lib uses 30 for those — they are quantwave's own surface, not a fidelity promise, and this is now recorded deliberately rather than falling out of a blanket default.

Fixed in the generator (scripts/gen_pyo3_plugins_py.py), not just its output: a per-function default table drives the emitted signatures, and a period-taking plugin missing from that table makes the generator fail rather than invent a uniform value. Regression cover lives in tests/python/test_ta_prefixed_defaults.py, anchored to a Pearson correlation and a TA-Lib BETA accumulation written out longhand in the test, plus guards that a bare call no longer returns the old 14-bar value.
BREAKING (Polars plugin surface): ta_atr, ta_natr and ta_trange now take close in the receiver, matching their non-prefixed siblings (quantwave-sww3).

These were generated with opaque positional parameters — ta_atr(self, in2, in3) — and forwarded as [self, in2, in3] to a plugin that consumes (high, low, close). The receiver therefore had to be high, while the sibling .ta.atr takes close. Writing the call the way its sibling reads permuted the inputs and returned a plausible wrong number with no error: on a 200-bar random walk, 1.252823 against a hand-computed Wilder RMA of 1.384819 — roughly 10% off, and data-dependent, so it never looked obviously broken. All three arguments are equal-length f64 columns, so nothing could raise.

This bit hardest exactly where it mattered least tolerably: you reach for a ta_-prefixed function specifically when you want TA-Lib-identical values, usually to reconcile against a chart.
```
# before — receiver had to be high
pl.col("high").ta.ta_atr("low", "close", timeperiod=14)
# now — same shape as .ta.atr
pl.col("close").ta.ta_atr("high", "low", timeperiod=14)
```
Parameters are now named high / low rather than in2 / in3, so a mis-ordered call is visible at the call site and keyword arguments work. ta_beta and ta_correl already matched their siblings' shape and are unchanged apart from in2 becoming other. The generator (scripts/gen_pyo3_plugins_py.py) was fixed, not just its output, so regenerating cannot reintroduce this. Regression cover lives in tests/python/test_ta_prefixed_arg_order.py, anchored to independently computed reference values rather than to current behaviour.
All 60 remaining candlestick patterns ported to native Rust — talib-rs has left the shipped dependency graph. The native_cdl! macro pushed every bar onto a 32-bar window and re-ran the full talib_rs::pattern::* batch function per bar, discarding all but the last value. All 60 are now O(1) streaming Next<T> implementations under quantwave_core::indicators::patterns, joining CDLDOJI and the 14 already-native single-bar patterns for 61 native candlestick patterns.

talib-rs moved from [dependencies] to [dev-dependencies] in quantwave-core, and the unused declaration was dropped from quantwave-py. It is retained solely as the #[cfg(test)] parity oracle and the benches/talib_comparison baseline, both of which stay. The "221 native indicators" claim is now literally true — nothing in the shipped wheel or crate delegates to C TA-Lib or to a third-party TA crate. Struct names are unchanged (CDLHAMMER, CDL2CROWS, …), so this is not a breaking change for any caller.

The dead native_cdl! macro and the eleven never-invoked talib_* macros were deleted from indicators/talib_wrapper.rs; the two live native macros (native_pointwise_1!, native_binary_2!) remain.
BREAKING (Rust API only): talib_rs::MaType replaced by a native quantwave_core::MaType (quantwave-wxii). Every matype field and constructor argument — APO, PPO, BBANDS, MACDEXT, MAVP, STOCH, STOCHF, STOCHRSI, and the corresponding quantwave-polars .ta() methods — now takes quantwave_core::MaType. The discriminants are unchanged (Sma = 0 … T3 = 8), so Python users are unaffected: matype= still takes the same integers and strings.

pub use talib_rs as talib; has been removed from quantwave-core. talib-rs is a parity oracle used from #[cfg(test)] and benches/, and a dependency in that role must not appear in a public signature. Rust callers substitute:
```
// before
use quantwave_core::talib::MaType;
// after
use quantwave_core::MaType;
```
MaType offers to_i32(), TryFrom<i32>, FromStr, Display, MaType::ALL, and from_u8_or_sma(). To cross to talib-rs in your own tests or benches, go through the integer code: talib_rs::MaType::try_from(m.to_i32()).
BREAKING: execution_delay now defaults to "next_bar" (T+1) instead of "same_bar" (T+0) (quantwave-zmjw). Affects .bt.backtest, .bt.backtest_with_report, .bt.backtest_metrics, .bt.portfolio_backtest and every other .bt entry point, the Python BacktestConfig, the Rust BacktestConfig::default() / ExecutionDelay::default(), and BtOptions::default() in quantwave-polars.

"same_bar" fills at the close of the very bar that produced the signal. Since signals are almost always derived from that same close (e.g. (rsi < 30) computed on bar t), the old default executed on information that only existed at the instant the bar ended — a look-ahead the live strategy never has, and one that flatters results systematically. On an identical signal frame over a rising series, same_bar entered at 100.5 where next_bar entered at 101.0.

Your existing backtests will report different — generally worse — numbers after upgrading. That difference is the look-ahead being removed, not a regression.

To restore the previous behaviour, pass it explicitly:
```
lf.bt.backtest(signal="signal", execution_delay="same_bar")
```
"same_bar" remains fully supported and is the correct choice when it genuinely describes your execution: you trade the closing auction, or your signal is built purely from data through bar t-1 so bar t's close is not an input. Otherwise, prefer the new default.

Fixed

calmar_ratio returned inf on a zero-drawdown run while the rest of the bundle returned NaN (quantwave-gz7d). quantwave-s3iu moved sortino_ratio and profit_factor to NaN for an empty denominator but left calmar_ratio = cagr / max_drawdown_pct out of scope, so a single-trade run reported two different conventions for the same condition — extended_metrics() gave sortino_ratio=nan, profit_factor=nan, calmar_ratio=inf. Calmar now returns NaN when max_drawdown_pct is zero with positive CAGR (and still 0.0 when CAGR is non-positive, the no-activity case).

This is not cosmetic. Walk-forward and sweep selection pick the argmax with a v > best_val comparison: inf > anything is True, so a degenerate variant that simply never lost would win the in-fold optimisation and be carried into the out-of-sample window, whereas NaN > anything is False and the undefined variant is skipped like a null. The grid selector is now the extracted, unit-tested select_best_objective, and both it and the TPE pool selector have explicit NaN-safety tests (including the all-undefined fallback: index 0 with -inf as the fold's train_metric).

The same sweep caught one more divergence: sharpe_ratio returned 0.0 when the return series had zero dispersion but a non-zero mean — a perfectly constant non-zero return is x/0, undefined, and 0.0 reads as a bad Sharpe for a run with no measurable risk. It now returns NaN, matching the sortino_ratio zero-downside-deviation branch. A genuinely flat series (mean ≈ 0), a series shorter than two observations, and an empty one all still return 0.0.

var_95 / cvar_95 were checked and deliberately left alone — they are quantiles, not ratios, with no denominator to be empty — as was benchmark, which stays None (rather than a NaN-filled dict) when alpha/beta are undefined. win_rate, total_return, cagr and avg_trade_pnl guard their divisors against the no-activity case and are unchanged. .metrics() remains the same 10 keys. - The Benchmarks Nightly workflow failed on every scheduled run: pyarrow was never installed (quantwave-ss7q). benchmarks/harness.py builds the pandas side of the memory-footprint comparison with polars .to_pandas(), which delegates to pyarrow — an optional polars dependency the workflow did not install. The build step passed, only the harness step died with ModuleNotFoundError: No module named pyarrow, so benchmarks/results/latest.json and the rendered docs/benchmarks.md had not been refreshed since the dependency drifted out.

pyarrow is now installed in .github/workflows/benchmarks-nightly.yml; the benchmark still compares against a genuine pandas frame, since it is what backs the "2-5x lower memory footprint vs pandas" claim. harness.py also gained a check_dependencies() startup check that names each missing module and prints the exact pip install line, so a missing dependency now fails in the first second rather than after data generation. --dry-run only requires polars and numpy, keeping scripts/quantwave_verify.sh runnable without the comparison stack.
Four candlestick patterns diverged from the TA-Lib reference in ways the random-walk parity tests never reached. Each substituted a different condition group for the reference's, and each was masked by a fixture that happened to satisfy both readings:
- CDLGAPSIDESIDEWHITE — had no EQUAL accumulator at all, substituting open[i] < close[i-1] for the reference's near-equal-opens test, and derived the signal's sign from candle colour instead of gap direction. The reference returns +100 for an upside gap and -100 for a downside gap with both candles white; the old code required a black pair for the bearish case, so it never emitted -100.
- CDLCONCEALBABYSWALL — tested the real-body gap between bars i-2 and i-3 instead of i-1 and i-2, required the first two marubozu shadows to be exactly zero rather than shorter than the SHADOW_VERY_SHORT average, and imposed a long-lower-shadow test on the 3rd bar that the reference does not have.
- CDL3STARSINSOUTH — the entire third-candle group was substituted: the reference's short-body, short-upper-shadow and short-lower-shadow tests were missing, replaced by body-containment within the 2nd candle and an exact-zero lower shadow, and the low/high containment was tested against the 2nd body rather than the 2nd bar's range.
- CDLBREAKAWAY — tracked the three-candle descent with close where the reference uses high/low, added a long-body requirement on the closing candle that the reference does not impose, constrained the 4th candle's colour (the reference leaves it free), and compared the final close against the 2nd/1st candle bodies rather than open[i-3] / close[i-4].
The fixtures in patterns/fixtures.rs for all four were rebuilt so they discriminate: each now contains blocks that fire under the reference semantics but not the old reading (and, where applicable, the reverse), separated by re-priming bars so the rolling candle averages return to known values. Each was verified to fail against the pre-fix implementation before the fix was applied. - MaStream silently substituted SMA for every matype except EMA (quantwave-ii0g). The enum carried only Sma and Ema variants and a catch-all arm mapped Wma/Dema/Tema/Trima/Kama/Mama/T3 to SMA, so APO::new(12, 26, MaType::Wma) returned an SMA-based APO with no error or warning. PPO, MACDEXT, STOCH, STOCHF and STOCHRSI were affected the same way. All nine families now dispatch to their native streaming implementation (MaType::Mama uses fastlimit = 0.5 / slowlimit = 0.05 and MaType::T3 uses vfactor = 0.7, matching C TA-Lib's ta_MA.c when reached through a matype argument). Values change for any non-SMA, non-EMA matype — the old numbers were the wrong algorithm. - BBANDS with a non-SMA matype was O(n) per bar and leaked memory (quantwave-3nyh). The non-SMA path appended every bar to an unbounded history and re-ran the batch bbands across the whole thing each tick — O(n²) over a series, in a struct whose SMA path is O(1). It is now genuinely incremental: the middle band comes from MaStream and the deviation is a two-pass sum of squared deviations over a rolling window of timeperiod values. Memory is bounded by timeperiod. Results are unchanged for MaType::Sma. - KAMA's streaming state retained the entire input history although only the last timeperiod + 1 samples are ever read; it now keeps a bounded ring buffer. The arithmetic, including summation order, is unchanged. - Shared-capital portfolio streaming ignored execution_delay (quantwave-zmjw). run_shared_capital_streaming_simulation passed the delay down to simulate_shared_capital, which discards it — the batch path pre-shifts signals per timestamp group, but the streaming path never did. Under SameBar (the old default) both paths agreed, so the bug was invisible; any caller who explicitly asked for next_bar silently got same-bar fills in streaming mode and broke batch↔streaming parity. The streaming path now applies the same per-timestamp-group shift as batch. - 68 indicators silently resolved to a streaming class instead of their batch function (quantwave-84cu). The generated TA registry introduced in 0.7.0 derived native batch symbol names with pascal_to_snake() (SuperTrend → super_trend), but the export_*! macros emit pub fn [<$name:lower>] (SuperTrend → supertrend). Every multi-word name missed; the 44 single-word ones (rsi, sma, atr) passed only because pascal_to_snake("Rsi") == "rsi". _resolve_ta_binding treated the miss as a fallback and returned native_streaming, so qw.supertrend was a class while qw.rsi was a function — with no error or warning. qw.supertrend(period=10, multiplier=3.0, high=…, low=…, close=…) again returns list[SuperTrendResult] as it did in 0.6; callers need no changes. - _resolve_ta_binding now raises ImportError when an entry declares a native_batch symbol the build does not export, rather than silently substituting the streaming class (whose calling convention differs). A native_batch of None still falls through to streaming/polars as before. - Corrected stale hand-written aliases in scripts/api_slug_aliases.json: fm_demodulator, fourier_series_model, my_rsi, precision_trend_analysis named non-existent snake_cased symbols; linreg, oc2, true_range declared batch exports that do not exist and now fall through to their polars methods; sr_monitor declared SrInteractionMonitor, a class never exported to Python.

Added

Batch↔streaming parity proptests parameterised over every MaType for MaStream, BBANDS, APO, PPO and MACDEXT, plus SMA/EMA/WMA/TRIMA coverage for STOCH / STOCHF. The pre-existing suite only ever constructed MaType::Sma, which is precisely why both bugs above survived unnoticed.
TA-Lib-parity streaming TalibWma, TalibTema and TalibMama (the last reusing the existing incremental Hilbert engine), each with its own parity proptest. The general-purpose WMA, TEMA and MAMA use different seeding and do not reproduce TA-Lib's values.
qw.trim_warmup() / qw.warmup_rows() (quantwave-4rsq). Indicator warmup is emitted as NaN, never null, so drop_nulls() / dropna() is a silent no-op on it and warmup rows flow into backtests and feature matrices unnoticed. qw.trim_warmup(frame, *specs, extra=0, strict=True) slices off the maximum warmup across every named indicator, keeping columns with different warmups row-aligned (unlike drop_nans(), which trims per column set). Accepts "rsi", ("rsi", {"period": 21}), {"rsi": {...}, "ema": {...}}, or an explicit int bar count, and works on DataFrame / LazyFrame / Series, including df.pipe(qw.trim_warmup, "rsi"). Unknown indicator names raise by default rather than silently trimming nothing.
.bt warmup warning (quantwave-4rsq). backtest, backtest_with_report, backtest_metrics, portfolio_backtest, walk_forward, monte_carlo and order_backtest now emit a quantwave.WarmupWarning when the signal or close column they receive starts with NaN/null rows. It is a warning, not an error — existing code keeps working — and is silenceable with warnings.filterwarnings("ignore", category=qw.WarmupWarning).
NaN-vs-null semantics documented prominently in the Python getting-started guide, the backtest quickstart, and the FAQ.
test_registry_native_symbols_resolve_against_build — asserts every declared native symbol exists in the compiled module. The prior test only checked a name was present, never that it resolved, so native_batch: "super_trend" passed cleanly. Plus a regression test that multi-word slugs bind as batch functions, not classes.

[0.7.0] - 2026-07-13

Added

Complete classic TA-Lib surface (quantwave.talib): 161 functions, up from 8 — RSI, MACD, SMA, EMA, ATR, ADX, BBANDS, STOCH, OBV, all 61 candlestick patterns, and the math/price transforms. The classic array-in/array-out API (talib.RSI(close, timeperiod=14), multi-output tuples, OHLC/candlestick inputs) delegates to the Polars .ta plugins, so values are the talib-rs-parity-tested Rust results. (quantwave-yp9a)
Top-level native symbol access restored: qw.FracDiff, qw.fracdiff, qw.rsi, qw.SuperTrend, … bind alongside the slug-based qw.ta namespace.

Changed

Unified the Python FFI on PyO3 (abi3), retiring uniffi. The indicator bindings, the Polars expression plugins, and the backtest engine are now a single PyO3 abi3-py39 extension in one crate (quantwave-py) producing one cdylib — a single maturin build yields one cp39-abi3 wheel with no wheel-merge step. (quantwave-5ipk.10, quantwave-6dgg)
Collapsed the three PyO3 crates (quantwave-python, quantwave-plugins, quantwave-backtest-py) into quantwave-py; deleted scripts/build_unified_wheel.py and consolidated to one pyproject.toml.

Fixed

Wheel tag / install correctness: the published wheel is now cp39-abi3 and installs correctly on CPython 3.9–3.13. 0.6.1 shipped a py3-none wheel bundling CPython-3.12-only extensions, which broke pip install on 3.9/3.10/3.11/3.13. (quantwave-9gek.1)

[0.6.0] - 2026-06-28

Added

Fractional differencing (FracDiff) (quantwave-wnd9): Prado-style stationary features; Rust Next<f64>, Polars lf.ta.frac_diff(), Python fracdiff()
HTML tear sheets (quantwave-0gi1): BacktestReport.to_html() / save_html() with equity, drawdown, and trade tables
Research loop (Tier 2): qw.build_feature_matrix(), lf.ta().features().recommended_matrix(), lf.bt.monte_carlo(), Rust .bt WFO-optimize + MC bootstrap
Product guardrails (Tier 1): scripts/quantwave_verify.sh, metadata drift gate, streamlined CI (verify → plugins → deploy-docs)
55 custom Polars expression plugins and 98 auto-generated pyo3-polars bindings for standard indicators
PA foundation: S/R Polars, confluence, geometric patterns with H&S neckline breakout
Streaming readiness (quantwave-h6xe) and Rust metadata codegen (quantwave-iqq7)
Plugin vs .ta guide, expanded regime user guide, comparison one-pager
Indicator doc SOA complete (quantwave-frq0): 220+ native pages under DOCUMENTATION_STANDARDS.md with PNG previews, doc drift script in verify
Full visual depth layer (p1k6): docs/generate_all_previews.py + standards lint rejecting placeholders

Changed

GitHub Actions consolidated from four workflows to CI + Release (v* → crates.io + PyPI)
Platform planning docs split into INDICATORS_SOA.md and BACKTEST_SOA.md

Fixed

CI: cargo-nextest, maturin venv, uniffi-bindgen==0.31.0 for verify job
Doc lint for fractional_differentiation.md (preview PNG + description depth)
Empty Python API Reference page (quantwave-rbz4)
Broken imports for quantwave >=0.4.1 on fresh installs without polars

[0.5.2] - 2026-05-31

Added (Python DX improvements)

Discovery API: quantwave.indicators() and quantwave.is_indicator(name).
Rich Metadata: quantwave.metadata(name) returning IndicatorMeta with params, data inputs, outputs, warmup_bars, category, etc.
Streaming lookup: quantwave.streaming_class(name).
Parity testing: quantwave.assert_parity() helper for verifying batch vs streaming bit-identical behavior.
warmup_bars(name, params) helper.
Namespace improvements: New quantwave.results, quantwave.options, and quantwave.talib submodules. Old top-level access now emits deprecation warnings.
Public exception base: quantwave.QuantwaveError.
__version__ properly exposed.
Linux arm64 (aarch64) wheels are now built and published.

Changed

Release workflow no longer hard-gates on docs build (docs issues can be fixed independently).

Documentation

Official Standards Published: Created docs/DOCUMENTATION_STANDARDS.md (v1.0, 2026-05-31 IST) under task quantwave-d2hk / epic p1k6. Defines mandatory enforceable template for all 223+ indicator pages: required sections (Visual Example, full batch+streaming+Polars Usage Examples, Edge Cases & Limitations, Sources), type-specific guidance (classic scalar / patterns / rich struct / Ehlers), good-vs-bad examples, tone/visual/cross-link rules, and 4-phase rollout.
Updated contributing.md (new indicator docs step) and appended full decision record + rationale (diagnosis of thin stubs vs. PA notebook quality) to DOCUMENTATION_DECISIONS.md.
Minor alignments in gallery.md.
This is the foundation for all future indicator documentation work and the planned xtask generator. See DOCUMENTATION_STANDARDS.md for the complete template and checklist.
Candle Standards Proof batch (p1k6 child, 2026-05-31 IST): 8 worst-duplication candlestick pages (doji.md + gravestone/dragonfly variants, harami.md + harami_cross, three_black_crows.md + three_white_soldiers.md, abandoned_baby.md) + engulfing.md enhancement fully rewritten to DOCUMENTATION_STANDARDS.md (mandatory visuals, 3-surface code, edges, authoritative TA-Lib+core sources, no Nison boilerplate). docs/gen_candle_previews.py extended (portable + 8+ generators); 11 professional PNGs produced in assets/candlestick-previews/. Cross-refs + full decision record in DOCUMENTATION_DECISIONS.md. Proves template + gens scale for Phase 1 rollout. See decisions file for files touched and bd tracking attempt details.
Ehlers DSP Phase 1 batch 2 (p1k6, 2026-05-31 IST): 5 high-value thin Ehlers DSP pages (ehlers_filter.md, reflex.md, ehlers_stochastic.md, ehlers_loops.md, ultimatesmoother.md) rewritten to full Ehlers/scalar STANDARDS conformance. Extended gen_indicator_previews.py (portable, pure-numpy core ports for the 5, CLI, professional DSP styling); 5 new PNG visuals generated with 2026-05-31 IST captions mapping directly to core .rs Next logic. 3-surface examples, Edge Cases, authoritative sources (exact core paths + Ehlers papers). Cross-refs + detailed decision record appended. Worktree clean for merge. See DOCUMENTATION_DECISIONS.md for complete list of files + checklist confirmation.

[0.5.1] - 2026-05-31

Fixed

Publishing completeness: Fixed workspace dependency configuration so that cargo publish succeeds for all internal crates (quantwave-core, quantwave-polars, quantwave-plugins, quantwave-backtest, quantwave). Internal crates now correctly declare version.workspace = true in [workspace.dependencies].
Release reliability: Added required build-docs job (export + mkdocs build --strict) to the release workflow. Release publishing now hard-gates on successful docs build. Removed all continue-on-error: true from publish steps — any failure is now fatal.
Docs build: Fixed filename collision (*.py.md landing pages conflicting with *.py notebooks) that was breaking main deploys. Renamed affected landing pages and cleaned up references + committed __pycache__.
Modernized cargo publish steps to use CARGO_REGISTRY_TOKEN environment variable (no more deprecated --token flag).

Changed

0.5.1 is the first complete, trustworthy release of the Backtest Engine v0.2 features (including quantwave-backtest crate on crates.io) plus the full Polars + Python package set.

[0.5.0] - 2026-05-30

Added

Backtest Engine v0.2 (major milestone):
Rich-Metadata Position Sizing: New PositionSizer trait + InitialRiskPositionSizer that directly consumes rich PA detector metadata (fraction_at_risk, pole_height_atr, strength, etc.) for dynamic, risk-aware sizing. Inspired by QF-Lib patterns. Includes SizingAdapter for seamless streaming Next<T> generators.
Pluggable Realistic Execution Models: Proper CommissionModel and SlippageModel traits with high-quality implementations, including SquareRootMarketImpactSlippage (volatility × √(volume/ADV)) and max_volume_share_limit support.
High-Fidelity Execution Simulator Mode: New execution path that applies the full sophisticated models while being driven by the exact same rich StrategySignal / PA struct stream as the fast vectorized path. Perfect for "pre-live" validation.
Professional Tearsheet & Reporting Layer: New BacktestTearsheet with PerformanceSummary, RiskMetrics, EnrichedTrade (carries full PA metadata for attribution), AttributionReport, to_markdown(), and Polars DataFrame export for Excel. Institutional-quality output.
Full batch + streaming Next<T> parity maintained across all new features.
Updated canonical examples and documentation demonstrating PA detector + rich metadata workflows.

Changed

Backtester is now production-grade ready for complex PA + ML strategies (Flags, H&S, Market Structure, etc.).

[0.4.0] - 2026-05-19

Added

Options India Analytics: Comprehensive suite for NSE options including Black-Scholes Greeks (Price, Delta, Gamma, Theta, Vega, Rho), Implied Volatility, and Chain Analytics (Max Pain, PCR, GEX, OI Zones, ATM Straddle, Synthetic Futures).
Polars Integration for Options: Full support for options_india as native Polars expressions with robust handling of column-or-value parameters.
NSE Utilities: Added nse_lot_size and moneyness helpers for the Indian market.

Fixed

Release Build: Resolved a critical 'maturin' conflict where tracked __init__.py files were being overwritten during the wheel build process.
Code Hygiene: Cleaned up all compiler warnings and unused imports across the entire workspace.

[0.3.0] - 2026-05-18

Added

Multi-Asset Regime Detection: Enhanced MultiAssetClusterer with rolling correlation structures and dispersion analysis to identify joint market states.
Advanced Conditioned Risk Metrics: Expanded regimes_conditioned_metrics in Polars to include Skewness, Kurtosis, and Sortino Ratio.
Polars Enhancements: Enabled moment and cum_agg features for vectorized higher-order statistics.

Fixed

Release Stability: Fixed workspace dependency alignment issues that caused CI failures in previous releases.
Compilation: Resolved method resolution errors for skew and kurtosis in Polars pipelines.

[0.2.0] - 2026-05-18

Added

Regime Detection Suite (quantwave::regimes):
- Volatility Clustering (Prakash et al. 2021) with online K-Means.
- Hidden Markov Models (Hamilton 1989) with Viterbi decoding.
- Gaussian Mixture Models (Two Sigma 2021) foundations.
- Changepoint Detection (PELT - Killick et al. 2012) for exact segmentation.
Polars integration for all regime detection tools.
Comprehensive documentation and guides for market state tools.