fynance 2.13.1

Pure-Python (Numba-accelerated) machine learning, econometrics and statistical tools for financial analysis and backtesting

Fynance

---

Pure-Python package (Numba-accelerated kernels) providing machine learning, econometric and statistical tools for financial analysis and backtesting of trading strategies.

Installation

pip install fynance

From source:

git clone https://github.com/ArthurBernard/Fynance.git
cd Fynance
pip install -e ".[dev]"

The build is pure-Python — there is no compile step (numerical kernels are Numba @njit, JIT-compiled on first call).

Architecture

A complete, layered ML/DL backtesting tool — data → features → signal → portfolio → backtest → metrics — composed through typing.Protocol seams. numpy is the lingua franca; PyTorch is confined to fynance.models. Each piece is usable standalone; fynance.strategy.Strategy is an optional orchestrator.

2.0 is a breaking release. See doc/MIGRATION-2.0.md for the import-path map (e.g. fynance.algorithms → fynance.portfolio, performance metrics → fynance.metrics).

Subpackages

Core fynance.core — PriceSeries value object (thin, numpy-backed) and the pipeline protocols (DataSource, FeatureTransform, SignalModel, Allocator, CostModel, Metric); executable conformance/causality checks (check_conforms, assert_causal) and duck-typed pandas/polars seams (from_pandas/to_pandas).

Data fynance.data — file adapters (load for CSV/Parquet → PriceSeries), alignment/resampling, no-lookahead temporal splits (train_test_split, walk_forward, combinatorial-purged combinatorial_purged_cv) and vendor-agnostic intraday session utilities (session_mask/session_id/split_sessions).

Features fynance.features — technical indicators (Bollinger, RSI, MACD, ROC, realized volatility, rolling skew/kurtosis/autocorr, …), OHLCV indicators (ATR, ADX, Williams %R, OBV, VWAP), a causal GARCH(1,1) conditional-volatility feature, momentums (SMA, EMA, WMA) and adaptive windows, NaN-aware cross-sectional operators (cs_rank/cs_zscore/cs_neutralize, …), pairwise rolling stats (roll_corr/roll_beta, cross_corr), fractional differentiation (fracdiff), AFML labeling (triple_barrier, meta_labels, uniqueness weights), scaling (incl. rolling rank), statistics, feature engineering (multi-resolution, Granger causality) and market-regime detection.

Metrics fynance.metrics — performance/evaluation metrics (Sharpe, Sortino, Calmar, drawdown, …) and a one-call summary; tail risk (VaR/CVaR/CDaR, tail dependence), benchmark-relative metrics (alpha, beta, tracking error, information ratio, capture), factor evaluation (quantile returns, rolling IC, IC decay), and turnover/exposure and per-trade analytics.

Signal fynance.signal — prediction → position mappers (sign, threshold, rank, vol-targeting) and a model+mapper pipeline.

Portfolio fynance.portfolio — allocation (ERC, HRP, IVP, MDP, MVP, plus risk-budgeting RBP) with an opt-in cov= seam over conditioned covariance estimators (Ledoit-Wolf, EWMA, factor, Marchenko-Pastur denoising); risk attribution, an exposure-constraints overlay (project_weights), rebalancing policies (calendar/band/turnover-cap, lot discretization) and sizing (fractional Kelly, single-series and book-level volatility targeting, transaction costs).

Backtest fynance.backtest — vectorized engine (backtest: positions + returns/prices + cost → BacktestResult), cost models (ProportionalCost, the non-linear MarketImpactCost, per-bar HoldingCost and CompositeCost stacking) and capacity analysis (capacity_curve, breakeven_fee).

Plot fynance.plot — composable matplotlib figures and one-call tearsheet and factor_tearsheet reports.

Models fynance.models — econometric models (MA, ARMA, ARMA-GARCH, plus GJR/EGARCH kernels) and PyTorch nets (MLP, RNN, GRU, LSTM, MultiHeadAttention, TCN, Transformer), a direction+magnitude stacking ensemble, RegimeMoE (regime-conditioned mixture-of-experts), objective-aligned training with cross-asset pretraining, distributional QuantileModel, uncertainty wrappers (DeepEnsemble, MCDropout), causal conformal intervals, purged walk-forward hyperparameter search, differentiable losses (Sharpe, Sortino, Calmar, Omega, directional, hybrid, pinball), and robust-training utilities.

Estimator fynance.estimator — volatility-model MLE: fit_volatility fits GARCH / GJR-GARCH / EGARCH with Gaussian or Student-t innovations, returning a VolatilityResult (AIC/BIC, conditional vol, forecast, simulate).

Strategy fynance.strategy — optional orchestrator composing the maillons end-to-end, with single-run and walk-forward evaluation.

Research fynance.research — data-agnostic experiment harness: Experiment (serializable run record), run_experiment (seeded, cost-aware, walk-forward), write_report (portable markdown + tearsheet PNG + notebook), synthetic data generators (gbm, regime_switching) and overfitting guards (permutation and block bootstrap, deflated Sharpe, PBO/CSCV, walk-forward MDA feature importance). Results are written only to a caller-provided output_dir — fynance never stores them itself.

Quick start

import numpy as np
import fynance as fy

# 1. Data — load a CSV/Parquet file, or build a PriceSeries directly
prices = fy.PriceSeries(100 * np.cumprod(1 + np.random.randn(750) * 0.01))

# 2. Compose a strategy: momentum feature -> position -> backtest with costs
strat = fy.Strategy(
    features=lambda p: np.sign(np.diff(p, prepend=p[0])),
    signal=lambda x: x,
    cost=fy.ProportionalCost(fee=0.0005),
)
result = strat.run(prices)

# 3. Evaluate and report
print(result.summary())     # Sharpe, Sortino, Calmar, max drawdown, ...
fig = fy.tearsheet(result)  # one-call performance report

See Notebooks/quickstart_v2.ipynb for the full runnable tour (data, features, walk-forward, reporting). An optional Streamlit playground ships under apps/playground/ (pip install -e ".[ui]" && streamlit run apps/playground/app.py).

Links

• PyPI: https://pypi.org/project/fynance/
• Documentation: https://fynance.readthedocs.io/en/latest/
• Source: https://github.com/ArthurBernard/Fynance
• Changelog: https://github.com/ArthurBernard/Fynance/blob/master/CHANGELOG.md