Long Arc

Single compiled Rust binary. No Python. No notebooks. No runtime dependencies. The same executable that runs backtests runs live trading—identical signal computation, identical risk controls, identical execution logic.

Every parameter is validated through walk-forward backtesting before deployment. Signal families are evaluated independently via information coefficient analysis with anchored expanding windows. No in-sample optimization reaches production without out-of-sample confirmation.

Eighteen independent signal families vote on position direction and magnitude. Each family captures a different market microstructure phenomenon—time-series momentum, cross-sectional momentum, pairs cointegration, funding rate arbitrage, structural breakouts, volume divergences, and mean reversion.

Signals are combined via z-score averaging of walk-forward-validated features. No tree-based model, no neural network. The combiner has five features and zero trainable parameters. Simplicity is the point: fewer degrees of freedom means less overfitting.

Position sizing follows fractional Kelly criterion with concentration caps. Correlated assets are clustered and their combined exposure is bounded. Net exposure is capped as a fraction of gross. Graduated drawdown scaling reduces all positions proportionally as the portfolio approaches its maximum drawdown threshold.

A setup-based entry filter requires multi-signal consensus before opening new positions. Early stop-losses cut underwater positions within the first 48 hours. A circuit breaker halts all new entries at the maximum drawdown level.