Walk-Forward Test an NSE Strategy

1. Freeze the strategy and its assumptions first

Open a supported NSE symbol and interval, then complete a backtest. Record the actual candle count and dates, capital, quantity, fees and slippage. Decide which parameters the code really reads. Pine grids use declared variable IDs, not input display labels or arbitrary execution-setting names.

Do not change the dataset between candidate evaluations. The optimizer uses the source, inputs, primary/secondary history and execution assumptions captured in the completed run.

2. Use a small interpretable grid

As an illustrative vectorized example, a strategy reading fast and slow can test fast=[10,20] and slow=[40,60,80]: six combinations. These are example values, not a profitable or recommended NSE strategy.

A 1,000-candle sample with 400 training candles, 100 test candles and three folds gives:

Offsets are zero-based, end exclusive. There are 18 training evaluations and three selected tests; 300 candles remain unused. Preview maps these offsets to the run's exact UTC dates. Do not apply this 1,000-candle example to sequential mode, whose source run is limited to 250 candles.

3. Choose training-only selection and warmup

Select highest training net P&L, which can include open mark-to-market, or lowest training max drawdown. Set a minimum completed-position count if you want to exclude near-inactive candidates. A maximum drawdown constraint is an account-currency amount, not a percentage.

Cold-start uses each segment's original configured warmup. Explicit past-only warmup uses preceding unscored captured candles, up to 5,000, with a fresh scored account. Missing preceding history shortens the warmup; positions and equity never carry across the boundary. Pine secondary warmup is unsupported.

Preview before running. It validates workload without executing the strategy. Optional measured time estimation executes one real candidate and is low-confidence, not a completion-time guarantee.

4. Run and read each phase separately

Start optimization explicitly and wait for its own terminal status. The baseline backtest can remain completed while optimization is still running. Cancellation, failure and skipped tests are distinct outcomes.

Read the winning training candidate, its eligibility and the following held-out result for every fold. Do not combine independently initialized folds and call the output a continuous portfolio return. Later training can include earlier test candles, so it is rolling validation rather than permanently unseen data.

Avoid repeatedly editing the grid after inspecting poor tests and then calling the same tests untouched. Keep both positive and failed attempts as research evidence.

5. Preserve the result without relabeling reused data

Save the stopped run with its current revision. A confirmed selected-parameter rerun creates another backtest on the full original sample, including reused tests. It is useful for inspection but is not new out-of-sample validation.

Detailed optimizer HTML generation reconstructs recorded evaluations and flags metric mismatches. Export HTML to retain that session-cached report. Small authored explanations belong to Research reports, not another optimizer run.

Use it with an AI assistant

Use tl-strategy-optimization and real commands backtesting.optimizePreview, backtesting.optimize, backtesting.optimizeMeasure, backtesting.optimizeRerun and backtesting.optimizeReport. Require explicit approval for operations that execute code and discover the actual Desktop tab's support first.

See the TradingLead command reference and AI prompt cookbook. Ask for training and test evidence separately, not a claim that the optimizer found a guaranteed profitable strategy.