- Systematic mid-frequency statistical-arbitrage on the Binance perpetual-futures market.
- Trades a broad cross-section of liquid USDT-margined perpetual contracts on a daily rebalancing cycle.
- Long and short positions sized to be market-neutral and dollar-balanced at portfolio level.
- Each contract enters the tradable universe only after passing a quarterly liquidity and continuity filter; the universe rolls forward through the backtest so that delisted or freshly-listed contracts are added and removed at the right point in time, preventing survivorship bias.
| Market | Binance USDT-margined perpetual futures |
| Frequency | Mid-frequency, daily rebalance |
| Style | Cross-sectional statistical arbitrage, market-neutral |
| Universe | 150 symbols, rolling quarterly refresh |
| Backtest window | 20210101 — 20260101 |
| Cost model | 5 bps fee + 2 bps slippage |
| Data sources | Binance Futures REST (price, volume), funding-rate history, open-interest archive, top-trader long/short ratio archive |
Three independent signals feed a portfolio combiner. All three share two construction invariants that are applied at every stage of the pipeline:
- Cross-sectional neutralisation. Every raw signal is residualised against a rolling market beta before any portfolio construction. The model therefore trades idiosyncratic dispersion rather than leveraged market exposure.
- Persistent long/short balance. The weight construction enforces approximately equal gross capital on the long and short sides each day, so the strategy retains its market-neutral profile across regime shifts rather than drifting net long in bull markets or net short in drawdowns.
- Cross-sectional preference for assets whose price trend is supported by an expansion in open interest — positions accumulating in the direction of the move.
- Funding-rate filter down-weights signals when the funding side disagrees with the price direction (a crowded-trade flag).
- Volatility-rank adjustment shrinks positions in unusually noisy names.
- Targets dislocations caused by liquidation cascades: an unusually large one-day drop in open interest is a forced-position flush that tends to mean-revert within a few days.
- Liquidation shock component flags the abnormal OI move.
- Regime score encompasses long/short ratio change to indicate shifts in market regime.
- Event-driven and naturally counter-cyclical to the momentum strategy.
- Explainable Boosting Machine trained walk-forward on the full factor panel.
- Non-parametric in its feature relationships — discovers piecewise shape functions and pairwise interactions rather than imposing them.
- Global model is trained with factors neutralised with respect to the regime separator, which is market volatility in our experiment.
- A mixture-of-experts extension trains a per-regime expert on the global model's residuals so that the residual structure inside each regime can be captured.
- The model is trained only with factors passing shape-health analysis.
- Daily mean-variance optimisation in strategy-return space.
- EMA smoothing as a post-process on the final signal.
All strategies use the same train / test split:
- In-sample (IS): 2021-01-01 → 2024-12-31 — used for parameter selection, walk-forward training (for EBM), and stability inspection.
- Out-of-sample (OOS): 2025-01-01 → 2025-12-31 — held out and evaluated only after the IS configuration is frozen.
Strategy parameters were selected by sparse grid search over the IS window, retaining only configurations that sit inside a manually-verified stable region of the parameter landscape (i.e. surrounded by neighbouring grid points with consistent IS performance). This deliberately avoids fitting to isolated peaks, biasing the selection toward parameter choices whose IS edge is robust to small perturbations.
- IS: 2021-01-01 → 2024-12-31 — used to choose
ts_lookback,sentiment_ma_window,half_life_decay, and the risk-control knobs (max_weight,min_active_symbols). - OOS: 2025-01-01 → 2025-12-31.
| Final return | 30.95 % |
| Annualised Sharpe | 0.65 |
| PSR (SR* = 0) | 0.9923 |
- IS: 2021-01-01 → 2024-12-31 — used to tune
lookback,smooth_lookback,funding_z_threshold, and the vol-rank adjustment. - OOS: 2025-01-01 → 2025-12-31.
| Final return | 65.69 % |
| Annualised Sharpe | 1.25 |
| PSR (SR* = 0) | 0.982 |
- IS (walk-forward training): 2021-01-01 → 2024-12-31 — every fold
re-trains on a trailing 252-day window inside the IS range and predicts
the next
retrain_freqdays. Used to fixtrain_window,retrain_freq,interactions,min_samples_leaf, and the MoE routing settings. - OOS: 2025-01-01 → 2025-12-31 — the model continues its walk-forward retrain cycle over this held-out year; reported metrics are computed only on these predictions.
- Top: per-fold feature importance.
- Bottom-left: in-sample vs out-of-sample information coefficient over time.
- Bottom-right: cumulative PnL of the EBM signal in isolation.
| Final return | 111.4% |
| Annualised Sharpe | 1.22 |
| PSR (SR* = 0) | 99.98 |
- Blends the three signals through the strategy-space mean-variance optimiser.
- IS: 2021-01-01 → 2024-12-31 — used to fit the strategy-space covariance shrinkage and EMA smoothing of the combined weight vector.
- OOS: 2025-01-01 → 2025-12-31 — reported metrics are computed on the held-out year only.
- The lower panel reports gross leverage and net exposure, long-short asset counts, the effective number of positions, daily turnover, the weight-magnitude histogram, and the most-traded symbols.
- Net exposure tracks zero and long-short asset counts move in lock-step — the persistent market-neutral and dollar-balanced behaviour described above.
| Total return | 97.7 % |
| Annualised Sharpe | 1.53 |
| PSR (SR* = 0) | .9974 |