← Back to the blogLeer en español
Strategy research

The overfitting problem in TradingView backtests

Why most TradingView strategy backtests are lying to you — and what to do about it before you risk a single dollar.

Updated: 10 October 2026

Donchian report on XAUUSD · VT Markets · 5 minutes, captured on 10 October 2026. Real market data; simulated trades. This screenshot does not establish overfitting.
Donchian report on XAUUSD · VT Markets · 5 minutes, captured on 10 October 2026. Real market data; simulated trades. This screenshot does not establish overfitting.

A strong historical result can be the outcome of a useful rule, a favorable period or repeated selection. The headline alone cannot tell you which. Overfitting becomes a concern when choices are tuned to the same observations later presented as evidence.

The search matters as much as the final script

Trying many configurations creates more opportunities to find an attractive result by chance. Reporting only the winner hides the selection process. Record the tested ranges, the number of trials and any discretionary changes to the rule.

The Probability of Backtest Overfitting examines this selection problem. There is no universal amount you can simply subtract from a reported Sharpe ratio to remove it.

Three ways a result becomes too specific

Parameter search: lengths and thresholds are selected because they fitted the same historical window used in the final report.

A narrow market sample: a trend-following rule is assessed only during a strong trend. That sample does not establish how it behaves in a range or a different volatility regime.

Manual inspection: a developer repeatedly checks historical charts while designing an indicator, then treats a strategy on those same charts as an independent test. Human decisions can fit a sample even without an automated optimizer.

These issues apply to Deepwick, TradingView and any other research environment. TradingView's strategy documentation also discusses overfitting and out-of-sample testing.

What additional tests can tell you

Shuffling a fixed collection of trades can explore sensitivity to their order. It does not reconstruct all the abandoned strategies or undo the process that selected the winner. The assumptions behind any resampling method determine what its output means.

Inspect nearby parameters during development. A broad region of similar behavior is worth investigating; it is not proof of a future edge. If you repeatedly inspect every configuration on the supposed test set, that set becomes part of selection too.

What you can do in Deepwick today

Deepwick includes an optimizer and a walk-forward analysis for saved backtests. Use optimization to ask how inputs change a result, then separate parameter selection from evaluation on later windows. Keep the cost assumptions visible; a realistic fill model does not correct selection bias.

Write your evaluation plan before starting, retain unsuccessful trials and reserve data you have not used to make design choices. Once you change the strategy after seeing a test, label the next run as another development iteration.

The next practical step is configuring walk-forward in Deepwick. Its result can challenge a hypothesis, but cannot certify future profitability.

backtestingoverfittingtradingview