pandas-ta

pandas-ta's implementations are verified against industry-standard calculations and match values produced by major charting platforms including TradingView for equivalent parameters. Manual implementations frequently introduce subtle errors — using a simple moving average instead of Wilder's smoothed average for RSI produces visually similar but numerically incorrect values. A custom implementation that diverges from standard calculations introduces a systematic discrepancy between the platform signals the manual strategy was designed around and the signals the bot actually generates, creating an undetected strategy translation error.

pandas-ta's rolling computation prevents look-ahead bias at the indicator level when applied correctly — each row's value uses only prior data. However, look-ahead bias can still be introduced in decision logic: if the backtest uses the indicator value from the current candle to make an entry decision about that same candle's close price — which would only be known at period end — the bias is in the decision logic, not the indicator calculation. Correct backtesting uses only indicator values from the previous completed candle to trigger entries on the current candle's open price.

pandas-ta works correctly with DataFrames of any size, including small ones containing only the lookback period required for the indicator. RSI with a 14-period lookback needs at least 14–15 candles to produce a valid value — passing a DataFrame with 100 recent candles produces accurate RSI on the most recent rows. For a live bot fetching the last 100 candles on each cycle, this is entirely sufficient. The library returns NaN values for initial rows where insufficient data exists, which signal logic must check before evaluating entry conditions on those rows.