ebt

Backtesting and evaluation toolkit for position-based trading strategies.

You bring a price frame (DatetimeIndex × assets) and a position frame on the same assets. ebt fits models walk-forward, turns predictions into positions, backtests them with an execution lag, reports key figures, and sweeps strategy variants into sub-portfolios. The hot loops (backtest, execution simulation, position combination, outlier detection) are Rust kernels built into the package as ebt_rust.

Features

• Walk-forward prediction (ebt.predict): retrain any fit/predict model on a rolling window with an embargo gap, for sklearn-style linear models and PyTorch nets.
• Feature and label helpers (ebt.features): multi-scale EMA filter banks, forward labels, OHLC resampling with causal timestamps, cross-sectional ranks.
• Backtest and key figures (ebt.simulate.kf): PnL from positions and prices with an integer or timedelta execution lag, Sharpe decomposition, turnover, holding time, Sharpe-by-lag, trend correlation. Wide (one column per asset) and stacked/long input.
• One-pager and ensemble sweeps (ebt.simulate.plot): one_pager evaluates one book and draws the standard figure; plot_ensemble evaluates a list of variants, splits them into sub-portfolios by config variable and summarises the sweep.
• Risk sizing and sanity checks: risk_adjust to a target annualised risk, risk_to_lots, cost-aware sizing (ebt.simulate.costaware), and a position_sanity_gate that flags broken books before anyone reads their Sharpe.
• Execution models: VWAP-ramped fills, spread-aware timing, trade suppression below a risk threshold; effective-spread estimators from daily OHLC.
• Futures tooling (ebt.futures): contract selection, rolls, Panama adjustment, blended rolls, interpolated and open-interest-weighted continuous series, position decomposition back into contracts.
• Streaming outlier detection (ebt.data.outlier): point-in-time detectors for trades, quotes and OHLCV bars.
• Parallel map (ebt.parallel): fork or lean-subprocess workers with zero-copy shared frames.

Install

Requires Python 3.9+ and a Rust toolchain (https://rustup.rs) to build the extension.

pip install ebt-backtest

Wheels are published for Linux x86_64 and Windows (Python 3.9+); other platforms, including macOS, build from the sdist with a Rust toolchain. From a checkout:

pip install .
# editable install (the top-level ebt_rust extension needs compat mode):
pip install -e . --config-settings editable_mode=compat

Extras: ebt[notebook] (IPython widgets for the tabbed sweep views), ebt[optimize] (cvxpy for the allocation optimiser), ebt[test].

Quickstart

import numpy as np, pandas as pd
import ebt
from ebt.simulate.linear import Ridge
from ebt.simulate.plot import one_pager, plot_ensemble
from ebt.simulate.tricks import risk_adjust

# synthetic 5-minute prices for four assets
rng = np.random.default_rng(0)
idx = pd.date_range('2022-01-03', '2023-12-29', freq='5min')
rets = 1e-3 * rng.standard_normal((len(idx), 4))
prices = pd.DataFrame(100 * np.exp(np.cumsum(rets, axis=0)), index=idx, columns=list('ABCD'))

# daily features and forward labels; a bar's timestamp is when its price is known
daily_ret = prices.resample('1D', label='right').last().dropna().pct_change()
features = ebt.apply_filters(daily_ret, lengths=[4, 16, 64]).replace([np.inf, -np.inf], np.nan).dropna()
labels = ebt.make_forward_labels(daily_ret, smoothing_length=5).reindex(features.index)

# walk-forward fit: retrain every 60 days on an expanding window, 10-day embargo
def walk_forward(model):
    out = ebt.predict(features, labels, models=[model], min_lookback='120D',
                      retrain_interval='60D', embargo='10D', progressbar=False)
    return risk_adjust(out[model], prices, shift='2min')   # scale to annualised risk 1.0

pos = walk_forward(Ridge(alpha=1.0))

# evaluate against the raw prices with a 2-minute execution lag
res = one_pager(pos, prices, shift='2min')
print(res['kf'])            # sharpe, sharpe exbias, turnover, holding(days), lptp, ...

# sweep variants and compare sub-portfolios per config variable
configs = [{'alpha': a} for a in (0.1, 1.0, 10.0)]
positions = [walk_forward(Ridge(alpha=c['alpha'])) for c in configs]
ens = plot_ensemble(positions, prices, configs_df=pd.DataFrame(configs), shift='2min')
print(ens['kfs_df'])        # one row per variant: config columns + key figures

Neural-net models work the same way: ebt.simulate.nn.NeuralNetModel(create_net=FeedForwardNet, n_hidden=[16], ...) trains with a masked MSE loss by default; pass create_loss_f for your own objective.

Conventions

• Timestamps mark when a value is known. Resample prices with label='right'; the bar stamped T holds prints in (T - freq, T]. make_ohlc and make_forward_labels follow this.
• Positions are frames on the same assets as the prices. NaN means "carry the previous position", 0 means flat. After risk_adjust a book has annualised risk 1.0, so key figures across books are comparable.
• Execution lag shift is an integer number of price bars or a timedelta ('2min', pd.Timedelta('5m')). A timedelta fills at the first price print at or after stamp + shift, so pass raw prices at a resolution finer than the signal.
• Stacked (long) input: backtest_stacked and friends accept one row per (time, contract) in pandas or Polars. Column names are auto-detected (ts / time / timestamp for time, series / symbol / contract for the contract, price / close / return for the value) and can be given explicitly with time_col=, contract_col=, pos_value_col=, price_value_col=.
• Futures tools take a columns={...} mapping (time, contract, expiry, price, weight) so any source schema works without renaming; outputs add a series column naming the continuous series ('FUT(front)', 'FUT(30d)', 'FUT(top3)', ...).

Key figures

Computed by one_pager and plot_ensemble ("kf" = key figure). Sharpe and risk use daily PnL.

Package map

Tests

pip install -e . --config-settings editable_mode=compat
python -m pytest tests
cd ebt_rust && cargo test --profile perf

Releasing

scripts/release.sh X.Y.Z tags a release and the GitHub workflow publishes it; see RELEASING.md.

License

MIT, see LICENSE.