Changelog#

All notable changes to this project are documented in this file.

[Unreleased]#

0.12.0 - 2026-07-23#

Added#

  • BayesianRegression: online Bayesian univariate regression that emits a causal one-step-ahead predictive mean and standard deviation plus the current slope and intercept. Normal-Inverse-Gamma posterior (noise learned online, Student-t predictive) with exponential forgetting (com/span/halflife/alpha) and a weak prior, so estimates and intervals are defined from the first sample.

  • Example notebook 16, "Supervised forecasting with forecast_pairs": build a leak-safe training set, fit a least-squares model, check it out of sample, and tie the prediction into a backtest.

Fixed#

  • forecast_pairs(count=..., dropna=True) now also drops rows whose target is NaN (not only feature-warmup rows), matching duration= mode, so it returns a clean training set.

Docs#

  • Documentation rewritten to the NumPy and scikit-learn voice. Removed the repeated batch-vs-live-stream property from per-operator docs and notebook closers (it is stated once, where streaming is the subject), the verification cells that only demonstrated equivalence, and the closing summary sections.

0.11.0 - 2026-07-22#

Added#

  • Backtesting: engines are renamed and reorganised into a Backtest<DataModel><OrderDef> grid. The data model (Price, OHLC, Trades, L1, L1Trades) names the market feed; the order definition (Target, Orders) names the strategy's output contract. Eight engines fill eight of the ten cells of the 5x2 matrix; BacktestPriceOrders and BacktestL1TradesTarget are intentionally not provided (see choosing_a_backtest_engine).

  • Target engines (BacktestPriceTarget, BacktestOHLCTarget, BacktestTradesTarget, BacktestL1Target): receive a scalar target position each event, compute the delta to the current position, and take liquidity to reach it. The target is clamped to the static [min_position, max_position] cap before sizing the order. BacktestOHLCTarget defers execution to the next bar's open (causal; no manual lag needed). BacktestTradesTarget and BacktestL1Target execute immediately against the current event.

  • Orders engines (BacktestOHLCOrders, BacktestTradesOrders, BacktestL1Orders, BacktestL1TradesOrders): receive a two-sided resting quote (bid_price, bid_size, ask_price, ask_size) each event plus the market data columns. Either or both sides can fill on the same event. A quote submitted already crossing the spread is a taker.

  • New engines completing the useful grid cells: BacktestOHLCTarget and BacktestOHLCOrders (replacing BacktestOHLC / BacktestOHLCMaker), BacktestTradesTarget and BacktestTradesOrders (replacing BacktestTrades / BacktestTradesMaker), BacktestL1Target and BacktestL1Orders (replacing BacktestL1), BacktestL1TradesOrders (replacing BacktestL1Trades), and BacktestPriceTarget (replacing BacktestSignal).

  • All engines accept min_position and max_position (default unbounded). Fills are capped by a three-way minimum: order size, participation limit (where applicable), and remaining room to the position bound.

  • MARKET/NaN/inf encoding shared across all engines: a finite price is a resting limit (maker); NaN is a side-agnostic market order (taker); +inf / screamer.MARKET is a market buy (never-fill sell); -inf is a market sell (never-fill buy).

  • Docs: choosing_a_backtest_engine grid overview page with the 5x2 matrix, order-definition interfaces, MARKET encoding table, and fill-cap rule.

  • SchmittTrigger gains an initial latch seed (0.0, 1.0, or NaN).

  • Delay(duration) stream op: re-stamp each event's index by a time offset (the time-based counterpart of Lag). Requires an explicit index; lossless, 1:1, no warmup.

  • screamer.supervised.forecast_pairs(X, y, count=|duration=): build a forecasting training set by lagging features to align with a future causal target. Returns (X_shifted, y, as_of); count= is event-based, duration= is time-based (uses Delay, needs an index). Fully causal (lags X, never leads y).

Changed (breaking)#

  • SchmittTrigger now seeds its latch with initial (default 0.0, low) instead of NaN, so a signal that starts inside the dead band reads low during warmup rather than NaN. Pass initial=nan to restore the previous undefined-until-first-crossing behavior; pass initial=1.0 to start high.

0.10.0 - 2026-07-18#

Added#

  • Risk statistics: RollingDownsideDeviation (downside semideviation, the Sortino denominator), RollingOmega (Omega ratio of gains to losses about a threshold), and RollingCVaR (historical Conditional Value-at-Risk / Expected Shortfall, the mean loss in the worst alpha tail; VaR is -RollingQuantile). Each with a reference page, a plotted example, and tests.

  • Backtesting: a suite of five causal C++ engines named by the market data they consume, all sharing one accounting core (detail::PnLAccount) and the [equity, pnl, position, cost] output schema. Fills follow one rule set: a trade or quote through your price fills the full remaining, a fill at your price is min(remaining, participation_ratio * available_size), and a marketable order fills fully with tick_size slippage. Resting orders fill at their limit price (maker); an order submitted already crossing is a taker.

    • BacktestSignal (2 inputs): marks a position signal to a price, with a fractional spread (crossing cost) and fee.

    • BacktestOHLC (6 inputs): a directional target-position strategy on OHLC bars, with market orders (fill at the open, crossing half the spread, paying taker_fee) and limit orders ("touch"/"breach" of the bar range, paying maker_fee); bars fill in full. Causal by design: the target decided on a bar's close is deferred and traded on the next bar, so no manual lag is needed.

    • BacktestTrades (4 inputs): a resting limit order against the trade tape; a through-print sweeps the full order, an at-print fills a participation_ratio share.

    • BacktestL1 (8 inputs): a two-sided maker against top-of-book quotes only. Fills are a documented heuristic: "breach" (default) fills on a through, "touch" adds a participation partial once per lock episode. Inventory cap, taker_fee, and tick_size for marketable orders.

    • BacktestL1Trades (10 inputs): the preferred maker engine. Quotes mark the position, the trade tape drives fills (each trade on its own event row, so no fill-versus-cancel ambiguity), a quote cross with no explaining trade is the run-over fallback.

    • BacktestReport (4 inputs, 6 outputs): the C++ node that turns an engine's [equity, pnl, position, cost] into the running report columns (dollar drawdown, cumulative cost, turnover, trade count, running max drawdown, and running Sharpe), so pure-C++ callers get the statistics too. The backtest_report helper wraps BacktestReport, labeling its columns and reading the last row into a summary (total PnL, max drawdown, cost, turnover, trades, Sharpe). It returns plain dicts of numpy arrays and needs no pandas. Reference pages with plotted examples, tests, and two demo notebooks (a signal on bars, and the event-driven engines on a real tape).

[0.9.0] - 2026-07-17#

More microstructure: order-flow toxicity, book pressure, and spread decomposition, plus a re-classified docs taxonomy.

Added#

  • Microstructure tranche 2, order-flow toxicity and book features:

    • VPIN (Easley-Lopez de Prado-O'Hara 2012): order-flow toxicity over a self-contained volume clock (equal-volume buckets, boundary-splitting).

    • MicroPrice (Stoikov 2018, first-order): the imbalance-weighted mid, a fair value that leans toward the thinner side of the book.

    • QueueImbalance: L1 book (queue) imbalance, a documented synonym of OFI applied to resting bid/ask sizes.

  • Microstructure tranche 3, order-book flow and spread decomposition:

    • ContOFI (Cont-Kukanov-Stoikov 2014): the canonical order-flow imbalance from L1 book events, distinct from the trade-flow OFI.

    • EffectiveSpread (2*|price - mid|) and RealizedSpread (the liquidity part kept after the price moves); their difference is the price-impact / adverse-selection component.

  • Each ships with a reference page, a plotted usage example, and tests.

  • Docs: the flat "Microstructure" function topic is split into four groups - Trade signing, Order-flow imbalance, Price impact & liquidity, and Order-flow arrivals - with "microstructure" kept as a search tag.

  • Docs: a third microstructure notebook (toxicity, book pressure, and spreads) covering VPIN, QueueImbalance, MicroPrice, ContOFI, and the effective/realized spread decomposition.

[0.8.0] - 2026-07-16#

Microstructure and order-flow operators, and the lazy Pipeline path moved into the C++ core.

Added#

  • Microstructure and order-flow operators, all implemented as C++ core nodes:

    • Trade signing: TickRuleSign, LeeReadySign, SignedVolume, and the bar-level BulkVolumeClassifier (BVC).

    • Order flow: OFI (order-flow imbalance) and RollingOrderImbalance.

    • Price impact and liquidity: RollingKyleLambda, EwKyleLambda, AmihudIlliquidity, RollSpread (Roll effective spread), and the Bouchaud Propagator.

    • Event intensity: HawkesIntensity, a self-exciting arrival-rate model.

  • Two demo notebooks driven by a small committed real-data slice (six hours of Deribit BTC- and ETH-perpetual trades): order flow and trade signing, and price impact and liquidity.

Internal#

  • The lazy (Python-iterator) Pipeline path now runs in the C++ core. A C++ driver merges the input feeds, drives the compiled graph, and runs the multi-output watermark as-of join, replacing the previous Python driver. The batch == lazy == graph invariant is now enforced by one implementation rather than by parallel ones agreeing. The k-way merge refills lazily, so it never reads further ahead than requested (safe for endless streams).

  • Eager combine_latest coalesces per index in C++ (the Python dedup is gone), and split runs in C++.

[0.7.0] - 2026-07-12#

The v2 API: one consistent call shape, streams as plain tuples, and pipelines.

Changed (breaking)#

  • Stream operators are now CamelCase config-first classes, called like the functors as Op(config)(data): Resample, Dropna, Select, Filter, CombineLatest, Merge. The lowercase function forms are removed from the public API.

  • The Stream class is removed. A stream is a plain (values, index) tuple; to_pandas / from_pandas convert to and from pandas.

  • Dag is renamed to Pipeline (from screamer import Pipeline).

  • Resample takes freq= (a time window) or count= (a number of arrivals), and agg= a functor or string. The every=, func=, and agg={dict} forms are gone; compose several reducers with CombineLatest instead.

  • replay and multi_resample are removed; compose the existing operators.

Added#

  • Comparison and logic operators: GreaterThan, LessThan, GreaterEqual, LessEqual, Equal, NotEqual, And, Or, Not, Where, IsNan, IsFinite. These build the masks the new Filter gate consumes.

  • Expanding whole-history statistics: ExpandingMean, ExpandingVar, ExpandingStd, ExpandingSkew, ExpandingKurt, ExpandingSlope, and the running ExpandingSum / ExpandingMax / ExpandingMin / ExpandingProd.

  • PosPart (max(x, 0)) and NegPart (max(-x, 0)).

  • OHLC bar aggregations for Resample: agg="ohlc", "ohlcv", "ohlcv2".

Internal#

  • All stream-operator compute moved into C++ (dropna, select, filter, and merge as graph nodes), so the Python bindings stay thin.

Fixed#

  • Clip vectorizes to SIMD min/max, about 2x faster; it had regressed to 2x slower than np.clip.

[0.6.0] - 2026-07-06#

Multi-stream and pipeline infrastructure.

Added#

  • The streams layer, for aligning and reshaping feeds that do not tick together: CombineLatest, Merge, Dropna, Filter, Select, split, and time- or count-based Resample.

  • Pipelines (then named Dag): wire operators into a reusable graph, run it in batch or live with identical results, serialize it to JSON, and visualize it as a text tree or Graphviz.

  • Despiking functors: RollingMedianAD, Hampel, ImpulseClip.

  • A topic taxonomy and unified frontmatter across all function reference pages.

[0.5.0] - 2026-05-21#

Added#

  • SchmittTrigger, a hysteresis comparator.

Changed#

  • The ignore NaN policy is applied consistently across the library: a NaN input is skipped and never corrupts a function's internal state.

[0.4.0] - 2026-05-20#

Changed (breaking - JSON consumers)#

  • screamer/data/help.json schema: the freeform body_markdown field has been removed and replaced with two structured fields:

    • details (string) - markdown prose, guaranteed to contain no fenced code blocks. Use this when rendering the description / math / notes.

    • examples (list of {language, caption, code}) - extracted code examples, one entry per ### Caption heading in the source markdown. {eval-rst} .. plotly:: directives are unwrapped to plain python.

    Consumers that read body_markdown must switch to details (and optionally render examples separately). No backwards-compatibility shim is provided.

Changed#

  • Function reference docs (docs/functions_*/<Name>.md) now follow a canonical layout: prose lives under H2 sub-headings (Description, Formula, …), examples live under a single ## Examples H2 with one ### Caption per example. The sphinx-rendered pages adopt the same structure.

[0.3.0] - 2026-05-11#

This release more than doubles the indicator surface (67 → 153 exposed classes) and closes six of the seven roadmap sections. Almost every new class is cross-validated against TA-Lib, pandas, scipy, or pandas-ta-classic to floating-point precision; documented divergences (EMA convention, Cutler vs Wilder RSI) are pinned by tests so future drift fires. Full test suite: 2126 passing.

Added#

Dispatcher infrastructure#

  • Plan E (N→M dispatch) completed. FunctorBase now supports any combination of multi-input + multi-output classes (1→1, N→1, 1→M, N→M). First consumers: Cart2Polar / Polar2Cart (2→2), KeltnerChannels (3→3), RollingLinearRegression (2→4).

  • New shared primitives: detail::MonotonicDeque<bool IsMax> (six+ classes share it), detail::WilderSmoother.

  • Documented multi-output shape rule output.shape == single_input.shape + (M,) with column-by-column pairing in docs/polymorphic_api.md.

Math#

  • Element-wise: Floor, Ceil, Round (banker's rounding), Square, Cube, Sin, Cos, Atan, Asin, Acos, Identity, Hypot (2→1), Atan2 (2→1).

  • Linear2(a, b, c) - two-input affine combination a*x + b*y + c, pairs with Sign/Relu/Sigmoid for compact composed expressions.

  • Cart2Polar / Polar2Cart - 2→2 coordinate conversions; first consumers of the N→M dispatcher.

Misc / data transforms#

  • CumSum, CumProd, CumMax, CumMin - O(1) running reductions matching numpy.

  • Diff2 - second-order finite difference (discrete second derivative).

  • Detrend(window) - x − rolling_mean(x).

  • Momentum(k) - alias of Diff(k) (TA-Lib's MOM).

Rolling-window statistics#

  • RollingMad, RollingIqr, RollingRange - composition of existing primitives or sharper variants.

  • RollingArgmin, RollingArgmax - window-offset of the rolling extremum (TA-Lib's MININDEX / MAXINDEX).

  • RollingRank, RollingPercentile - pandas-style position metrics with average tie rule.

Exponentially-weighted statistics#

  • EwCov, EwCorr, EwBeta - 2-input pair statistics. Matches pandas ewm(adjust=True).cov / .corr bit-exactly. EwBeta follows the CAPM (target, regressor) convention.

Moving averages#

  • WMA - linearly-weighted moving average, O(1) per step via the identity W[t] − W[t−1] = w·x[t] − S[t−1].

  • DEMA, TEMA - Mulloy's double/triple EMA compositions.

  • TRIMA - triangular MA (SMA(SMA(x))).

  • HullMA - WMA(2·WMA(n/2) − WMA(n), √n).

  • KAMA - Kaufman Adaptive MA with O(1) per step; matches TA-Lib bit-exactly.

Momentum / oscillators#

  • MACD (1→3), WilliamsR (3→1), Stoch (3→2, fast and slow via smooth_k), StochRSI (1→2), TRIX, BOP (4→1), CCI (3→1), UltimateOscillator (3→1), ADX (3→3 returning +DI / -DI / ADX).

  • ROC, ROCP, ROCR - TA-Lib rate-of-change family.

  • RollingRSI default changed to Wilder's smoothing (matches TA-Lib and pandas-ta); method="cutler" preserves the old Cutler form. Earlier versions diverged from TA-Lib's RSI by ~11 RSI points -- now bit-exact.

Volatility / range#

  • Range-based volatility quartet: Parkinson, Garman-Klass, Rogers-Satchell, Yang-Zhang. Each ships in Var and Vol variants, with rolling and EW smoothing (for the first three) -- 14 classes total.

  • TrueRange, ATR(window), NATR(window) - Wilder's bar-aware volatility family.

  • DonchianChannels (2→3), KeltnerChannels (3→3) - channel indicators.

Volume-aware#

  • RollingVWAP, OBV, AD, ADOSC, MFI - first volume-aware primitives in screamer; all bit-exact to TA-Lib counterparts (except ADOSC, which inherits the documented EMA-convention divergence).

Performance / risk#

  • Drawdown, MaxDrawdown, RollingMaxDrawdown, RollingSharpe(window, periods_per_year), RollingSortino(window, ppy, target), RollingInfoRatio(window, ppy), RollingCalmar(window, ppy), RollingHitRate(window) - backtest-evaluation metrics. None of these are in TA-Lib; they're a real differentiator for screamer in trading pipelines.

Statistical / regression#

  • RollingAlpha - companion intercept to RollingBeta.

  • RollingResidualStd - std of the per-bar RollingSpread.

  • RollingLinearRegression (2→4) - full OLS fit returning (slope, intercept, r², stderr). First 2→4 consumer of the N→M dispatcher. stderr is the RMSE of residuals (standard error of estimate, not slope-stderr).

  • RollingTSF - TA-Lib's Time-Series Forecast (regression vs time projected one step ahead), bit-exact to talib.TSF.

  • RollingHurst(window, min_scale=4, method='rs') - rolling Hurst exponent via Anis-Lloyd corrected rescaled-range analysis at dyadic scales. Bit-exact to the reference Python implementation; ~0.5 on white noise, >0.5 on integrated processes.

Signal processing#

  • ButterHighpass, ButterBandpass, ButterBandstop - high-pass, band-pass, band-stop Butterworth IIR filters. Added the underlying lp2hp_zpk / lp2bp_zpk / lp2bs_zpk ZPK transformations so future Bessel/Cheby/Elliptic families also get all four btypes once their prototypes are written.

  • MovingAverage(taps) - FIR with arbitrary user-supplied taps (pair with np.hamming / np.kaiser / scipy.signal.firwin).

  • KalmanFilter(process_var, observation_var) - scalar 1-D Kalman, O(1) per step.

Validation#

  • New tests/test_third_party_alignment.py runs against TA-Lib and pandas-ta-classic under the new optional validation install group. Tests are skipped gracefully when the libraries are unavailable.

  • New docs/conventions.md documents the few deliberate divergences (EMA adjust=True vs TA-Lib's adjust=False + SMA seed; ddof=1 vs ddof=0 in RollingStd; Wilder vs Cutler RSI). Each divergence is asserted in the expected direction so future drift in either screamer or the third-party library trips the test.

Changed#

  • RollingRSI default is now Wilder smoothing (was Cutler); method="cutler" preserves the previous behaviour.

  • Pre-existing pending items now part of this release: general- order Butterworth filter, RollingOU, refactored devtools.

Version v0.1.46 (2024-11-02)#

Added#

  • RollingSigmaClip

  • Relu

  • Elu

  • Selu

  • Sigmoid

  • Tanh

  • Softsign

  • Linear

  • Power

Version v0.1.35 (2024-11-01)#

  • Improved documentation

Version v0.1.34 (2024-10-31)#

Version v0.1.33 (2024-10-31)#

Added#

Data handeling#

  • fillna

  • ffill

  • clip

Math#

  • Abs

  • Sign

  • Exp

  • Log

  • Sqrt

  • Erf

  • Erfc

Simple transforms#

  • Diff

  • Lag

  • Return

  • LogReturn

Rolling window functions#

  • rolling std

  • rolling skew

  • rolling kurtosis

  • rolling zscore

  • rolling min

  • rolling max

  • rolling median

  • rolling quantile

  • rolling rms

  • rolling poly1, 1rst order polynomial fit

  • rolling poly2, 2nd order polynomial fit

Exponentiually weighted functions#

  • EwMean

  • EwStd

  • EwVar

  • EwSkew

  • EwKurt

  • EwRms

Filters#

  • 2nd order Butterworth

Interface#

  • support for iterator / generator processing

Fixed#

  • Fixed incorrect results when applying transforms to a view on a numpy array.

Changed#

  • removed the transform member functions

Version v0.1.32 (2024-10-20)#

Added#

  • Differences, Simple Return, Log Return, Rolling Sum, Simple Moving Average

Changed#

  • removed initial_value from the constructor, we (for now) return NaN values when we cant resolve indicators.

Version v0.1.31 (2024-10-20)#

Added#

  • The indicator.transform() member functions can now transform multi-dimensional numpy arrays.

  • Added documentation.