How to test a machine-learning indicator

Treat a machine-learning indicator as a testable idea. Its name, chart colors and historical fit are not evidence that it will improve your workflow.

Write the experiment first

1. Define the decision the indicator is meant to improve.
2. Choose a simpler baseline with the same role.
3. Fix the symbol, timeframe and decision rules.
4. Use one period for configuration and a separate later period for evaluation.
5. Keep a record of every configuration you tried.

Do not select a “winning” version using the evaluation period and then describe that same period as an independent test.

Check the simulation assumptions

If you convert signals into a strategy, inspect fill assumptions, fees, slippage and available information at signal time. TradingView explains these concerns in its strategy-testing documentation.

Record missing signals and live behavior as carefully as attractive examples. State how much evidence you have and which market conditions the sample does not cover.

We have not independently backtested the catalog's machine-learning scripts. Use the library to find candidates, not to infer performance.