A model earns a further look only after it passes the held back test with costs

A model earns a further look only after it passes the held back test with costsDiagram with 6 boxes. Write the label as a net trade result, after costs to Keep only features that were known at the time of the trade; Keep only features that were known at the time of the trade to Split by time: fit on the past, tune on a middle period; Split by time: fit on the past, tune on a middle period to Count every model and feature tried, not only the best; Count every model and feature tried, not only the best to Run the held back period once, with costs, and log it; Run the held back period once, with costs, and log it to Repeat on fresh data before any claim is made.Write the label as a nettrade result, after costsKeep only features that wereknown at the time of thetradeSplit by time: fit on thepast, tune on a middle periodCount every model and featuretried, not only the bestRun the held back periodonce, with costs, and log itRepeat on fresh data beforeany claim is made
The order of checks that a machine learning trading idea passes before its score is read. A map of method, not a result. Source: RupeeCase method map; SEBI study on F&O outcomes (Aug 2026). As of 16 Oct 2026.

Machine learning is a set of methods that fit a function to data. In trading the function maps features, such as recent returns, volume and open interest changes, to a forecast or a trade decision. The method is rarely the weak point. The label, the features and the test are.

Three things decide whether a model deserves a further test. The label must be what a trade could earn after costs, not an arbitrary next day direction. The features must be known at the moment of the trade, which rules out any field that is revised later. And the test must run on data the model never saw while it was built.

Failures tend to come from a few repeated mistakes. Overlapping labels make the sample look bigger than it is. Many variants tried on one history make the best one look lucky. Survivorship and look ahead bias flatter the result. A model that loses in a fresh period is telling the truth, even when its earlier score was high.

The flow below is a method map, not a result. It shows the order in which a model passes through checks before its score is worth reading. Each arrow is a stage where a model can fail quietly.

SEBI's Aug 2026 study of all individual equity derivatives traders in FY26 is a reminder of the stakes. It found 87.7% lost money after costs, and fewer lost before them. The gap between the two is the traders for whom costs decided the sign, and a model that ignores costs has not tested what a trader would keep.

Before you trade. Derivatives trading carries a high risk of loss. SEBI's study "Profitability of Individual Traders in the Equity Derivatives Segment (FY25 to FY26)" (August 2026) found that 87.7% of individual traders in equity F&O made net losses in FY26, with aggregate net losses of Rs 91,685 crore. Read the SEBI study