Historical outcomes are not executable quotes
A backtest may identify the settled outcome but real profitability depends on the order price, available depth, and actual fills. Final win/loss labels cannot establish that every historical trade was executable at an ideal price.
Data time and resolution sources matter
Binance market data, Chainlink references, and Polymarket order books serve different roles. Verify the market rule, reference source, time window, decision timestamp, and observed arrival time independently.
Causal backtests can still overfit
Avoiding look-ahead does not eliminate model-selection bias from repeatedly testing on the same history. Inspect stability across periods and market environments, and continue with genuine forward observation where possible.
Does production reproduce research trade by trade?
Matching aggregate win rates or scores is insufficient. Compare individual events, skip/stop states, readiness, and boundary decisions between production and replay.
Execution has its own failure modes
Partial fills, rejected orders, latency, liquidity gaps, fees, and account-state issues can make real trading diverge from research. Monitoring and recovery belong inside the trading system.
The takeaway
Backtesting supports limited historical claims, not guaranteed future returns or real executability. Data, validation, prices, and production behavior should be checked separately.
For technical education only, not financial advice. Confirm exact features and requirements in the latest software release and project documentation.