⚡
Built with agentic AI. Development on this project runs through multi-agent Claude Code
workflows — developer → independent-reviewer cycles with evidence gates, batched into unattended runs that
execute overnight and resume from committed ledgers. How the agent system works →
Project overview
Historical OHLCV lives in SQLite; a backtest engine computes market structure (pivots, unbroken levels, session high/low for NY, London and Asia) over it, and the same code path drives live trading: bars stream in from Rithmic, the chart updates in place, and bracket orders route to Rithmic or Sierra Chart when a strategy signals. The desktop chart offers timeframes, indicators, drawing tools and a "go live" mode; strategy results export to Excel with per-trade statistics.
Market-structure engine
- An "unbroken level" engine tracks swing pivots per direction and timeframe, maintains the levels price hasn't broken, and emits structure/retest events.
- A streaming bar aggregator builds arbitrary timeframes (including renko) tick-by-tick from one source stream — identical for backtests and live.
- Trade accounting supports partial exits, running max-runup/max-drawdown, and per-trade feature export.
Leakage-aware ML trade filtering
- XGBoost classifiers on "did this trade profit", with splits grouped by futures contract and a whole held-out contract as out-of-time data.
- The probability cutoff is chosen on validation only; calibration curves are part of every run.
- Before a feature may exist: mutual information, a stump AUC with bootstrap 95% CI, and a label-permutation p-value. Feature manifests are serialized next to each model so live inference can never silently reorder columns.
Dual-broker live execution
- Bracket orders (entry + SL + TP as OCO) submitted through Sierra Chart's SCBridge, or through Rithmic's R|Protocol directly — protobuf over TLS WebSocket.
- The live consumer slices only genuinely-new bars from each overlapping realtime frame with a binary-search watermark — idempotent under duplicate pushes by construction.
- Backtest entries are code-generated into both a Pine Script indicator and a C++ study, so simulated fills can be visually reconciled on TradingView and Sierra charts.
Native C++ orderflow studies
- ACSIL studies read Sierra Chart's binary intraday file directly with shared buffered I/O — fixing a documented production deadlock that starved the live tick writer.
- A bar is only finalized once the file provably contains the next bar's open, so a lagging disk flush can never truncate it.
- Volume-anomaly detection scores bars against context-adaptive baselines (the conditioning stays private); every bar is replayed tick-by-tick and the first signal print is latched permanently, so recalculation reproduces exactly what a live watcher saw.