TradingView to 2.45-Billion-System GPU Backtest Pipeline
Extract full chart history from TradingView, sweep billions of indicator combinations on CUDA, and deliver only robust settings as Pine.
Chart-driven history extraction over CDP, a 2,449-variant / 21-family indicator library combined into 2.45 billion three-indicator systems, out-of-sample selection, null-data tests and Monte Carlo, packaged as a reusable skill.
- Full history extraction over CDP with symbol/timeframe restored afterwards
- 2.45B systems in ~2.7 minutes on a single GPU
- Null-data sweep: shuffled gold beat real gold 4 of 5 times
- Delivered Pine matched TradingView's tester within 0.2%
Clients ask for "backtest all available data on this symbol" while their chart is sitting on something else. This pipeline answers that end to end. It drives TradingView Desktop over the DevTools Protocol to page back through history until the server reports end of data, reads every bar, then restores the original symbol and timeframe, with zero npm dependencies. On this account that meant 21,370 bars of 5-minute gold or SPY daily back to 1993.
The engine evaluates a library of 2,449 indicator variants across 21 families as 2.45 billion unique three-indicator systems on an RTX GPU, about 2.7 minutes for 15,000 bars. Selection is deliberately pessimistic: systems are ranked on their worse half, plateau neighbours are counted, a CPU oracle re-runs the winner, and a null test rebuilds the series from its own shuffled returns and re-runs the full sweep. On gold, four of five shuffled datasets beat the real in-sample leader, which is the cheapest rigorous proof that an in-sample leaderboard maximum is a data-mining artefact. A trade-bootstrap Monte Carlo then reports drawdown percentiles rather than a single backtest number.
The skill's own first run caught four real scoring bugs before shipping, and the final Pine v6 strategy matched the TradingView Strategy Tester within 0.2 percent.
How it works
- Driver connects to TradingView Desktop and requests more data until end-of-data
- Bars are written to CSV and integrity-checked
- CUDA engine sweeps all three-indicator combinations with costs applied
- Robust picks selected on worse-half performance, plateau and CPU oracle agreement
- Null-data and Monte Carlo tests quantify luck and drawdown
- Pine v6 strategy generated and verified in the Strategy Tester
Software and automation are infrastructure, not financial advice. No performance guarantees. Every live system needs staged testing, risk limits, and owner approval.