Multi-year evaluation of MAE-bounded mechanical rules

The first window stayed unfinished

The evaluated procedure was a mechanical trading system: a complete, testable procedure that turned rule inputs and market state into entry, exit, or abstention without discretionary rewriting during the test. It combined a small set of mechanical entry and exit rules with two stops.

Those stops were chosen with maximum-adverse-excursion analysis, a trade-level map of how far price moved against the position, so large adverse moves were cut while winning trades still had room.

The first test window produced fewer than 30 trades, treated as below a conventional reliability threshold. Further calendar windows were prioritized over refining the logic on that short sample.

Stops stayed as a pre-placed bound

On a second year of the same bond-futures market, the procedure was rerun with and without the MAE-selected stops. The with-stops version was kept after that comparison.

A 700-unit stop, later observed as 765 after slippage and costs, truncated most large and many mid-sized losses, except on two gap days when the stop could not be filled as placed.

Stops were judged against a per-trade bound of 2 percent of capital. That check addressed individual losses but not consecutive-loss drawdown, the equity path created by a run of losers, which a per-trade stop does not measure by itself.

Filters were not retuned for later noise

Indicator plots were treated as untrustworthy because of display artifacts, so trades were checked against the numerical indicator values rather than the charts alone.

A filter threshold is a small logic gate meant to ignore one-day indicator noise, and it should come from inspecting raw behavior, not from searching for a later-sample profit maximum. One-day indicator anomalies of about 0.003 and 0.016 after an existing 0.005 filter were not used as a reason to retune that filter, because the 0.005 value had already been chosen by inspecting raw no-stop results.

A whipsaw cluster is a patch of too-frequent trades that suggests the rules are flipping in quiet or two-sided action. Clusters of extra trades and possible whipsaw appeared in more than one year and were logged as a candidate filter re-estimate only after those years could be inspected together.

A multi-day buy that matched the indicators but still failed in a decline was left unchanged. Rewriting rules to erase a few misses can also distort trades that already fit the concept.

Later years still left work open

Robustness testing means rerunning the same mechanical procedure on additional windows of similar length but different activity to see which problems repeat before any rule is changed. After that walk through several additional years, unfinished work still included consecutive-loss drawdown, pyramiding, and a pooled filter re-estimate.

Period transfer asks whether the same rules, stops, and filters still produce a coherent trade set when the calendar window changes but the instrument does not. That was the evaluation that was actually done. Transfer to other instruments was deferred because the rules were built on one market's observed behavior. Instrument specificity is the limit that this behavior should not be assumed to apply to unrelated markets until it is re-established.