Forecasting
Backtesting Done Right: Walk-Forward Testing Without Look-Ahead Bias
MarketLens Editorial · 12 Oct 2026
Backtesting asks: if we had used this model in the past, how wrong would it have been?
The leakage trap
Randomly splitting time-series data lets a model "see" the future. Results look great and mean nothing.
Walk-forward testing
MarketLens repeatedly picks a past date (the origin), fits each model only on data up to that date, forecasts h sessions ahead, and compares with the real price later observed. The origin then moves forward and the process repeats.
Metrics we report
- MAE — average absolute error in price terms.
- RMSE — penalises large misses more heavily.
- Directional accuracy — how often the forecast direction matched reality.
We always compare against the naive random-walk model. If a fancy model cannot beat "tomorrow equals today", you should know.
Risk notice: Forecasts are uncertain statistical estimates. Scenarios are algorithmic outputs, not investment advice. EatoRush MarketLens is not SEBI-registered. See the Risk Disclaimer.
Related articles
- Backtesting without look-ahead bias
Why every MarketLens backtest only uses data that existed at each forecast date.
- Why Every Forecast Needs an Uncertainty Range
A single number hides how wrong a forecast could be. Here is how MarketLens measures that.