A calendar for walk-forward robustness and Monte Carlo evaluation

Evaluation as a calendar

Editorial view: system evaluation is a calendar, not a trophy hunt. Freeze a two-year in-sample fit, survive a robustness cull, walk the next six months, then resample that unseen path before the next retest date.

A rules-based evaluation procedure first chooses the in-sample and out-of-sample windows, then searches parameters only on the in-sample side. In-sample is the historical window reserved for parameter search. Out-of-sample is the unseen window reserved for evaluation.

Windows come before the search

The window design uses two-year in-sample blocks and the next six months as out-of-sample, an 80/20 split that implies reoptimization every six months.

After an in-sample choice, the procedure builds a walk-forward equity path on unseen data, then uses Monte Carlo simulation to resample that path, sketch the next six-month equity range, and repeat the cycle six months later.

The rule is reduced before the sweep

A prior three-argument regression rule was reduced to two lookbacks after the slope term showed the weakest contribution. The long-horizon filter and dynamic sizing were removed before retesting.

Two integer rule inputs are swept from 5 through 30, producing 676 combinations that are filtered rather than ranked by peak net profit.

Robustness testing filters combinations

Robustness testing requires beating a period hurdle, then keeps only win-rate and win/loss ratios between the 20th and 80th percentiles. That step removes about 90 percent of combinations before a profit-factor cap. Profit factor is gross profit divided by gross loss.

A 1,000-run Monte Carlo simulation of 126 independent trades approximates profit factor from win rate and average win/loss so combinations implied above 2 can be dropped.

Remaining pairs are judged by whether they sit on a high, locally flat net-profit surface in two-parameter space instead of on an isolated peak.

Walk-forward analysis stitches the unseen path

Each retained pair is applied only to its matching unseen six-month window, and those daily equity changes are stitched into one out-of-sample path.

Monte Carlo simulation sketches the next range

Forward paths resample those out-of-sample daily returns with replacement over 126 sessions, 1,000 times, under an independence assumption later flagged for an autocorrelation check.

Editorial view: the resampled range is a sketch of the next six months on that unseen path, not a prize for the peak in-sample pair. The next retest date is part of the same calendar.