Walk-forward reoptimization as a system design gate

The laboratory rule set

The illustrated mechanical trading system exposed three tunable inputs: a long-entry lookback length, a long-entry fraction of a recent price range, and a clock time used to flatten shorts. Those inputs sat inside a fully specified procedure, not a loose collection of chart ideas.

Long entries were next-bar limit orders taken a fraction of a recent high-to-low range below the session high. Short entries were next-bar market orders. Shorts were covered at market after a specified clock time except at the session close, and a fixed initial stop of 600 currency units was attached to open risk.

The procedure allowed at most one entry per session except at the session close. Most bars were therefore an explicit abstention rather than a forced trade.

The first static window

The first evaluation window was a six-month static-parameter test. It was presented only as a foundation for later robustness work, not as a finished verdict on the rule set.

System optimization over the three inputs could produce a fitted window. That search remained unfinished laboratory work until the chosen values were tested after they left the fitting window.

Reestimation on a later cycle

Walk-forward analysis periodically re-estimated the same inputs to test whether the rule set remained usable or was only numerical debris from one fitting run. The procedure itself did not change: entry, exit, and abstention rules stayed one mechanical trading system.

Inputs were reoptimized on a cycle of about 700 price bars on an intraday gold-market series instead of being left frozen after the original development window. That bar count is the reoptimization cycle: after it elapses, the procedure selects a new input set before the next unseen test.

The five-category scorecard

The walk-forward scorecard judged the candidate on five categories: overall profitability, walk-forward efficiency, consistency of profits, distribution of profits, and maximum drawdown. Walk-forward efficiency compares unseen-segment results with the fitting windows used to choose the inputs.

An out-of-sample stress test evaluates data that were not used to choose the current inputs. In this workflow it functions as a pass-or-fail design filter, not as a second look at the original fitting run.

Leaving the modeling stage

The write-up treated continual reoptimization, stress testing, and out-of-sample forward tests as the process that decides whether a mechanical trading system can leave the modeling stage. A single static backtest was not that process.

The archive facts describe this historical workflow. They do not turn the laboratory rule set into a finished trading recommendation.