Session-window trims for mechanical day-system evaluation

A three-layer evaluation

Editorial framing: treat evaluation as three layers that stay separate. First specify one mechanical trading system so entry, exit, and the end-of-day flatten are tested as a single procedure on rule inputs, market state, and execution assumptions. Next use session analysis to choose which minutes of the order-life cycle enter the series, after weighing auction imbalance, missing cash references, spread, volume, and trading-cost conditions. Then robustness-test the identical rules after changing only the session window.

Editorial note: a moderate trim and an over-trim are shown side by side so students can see when data-reduction stops cleaning session-edge noise and starts deleting the path the signals need. The archive does not report that the first trim improved results. It reports that the tighter window produced weaker evaluation metrics.

Session-edge noise as a series filter

Around-the-clock equity-index mini futures concentrate most activity in the regular cash session. The final 15 minutes after the cash close lack updated cash-index and major-stock prices while absorbing piled-up closing orders. The first 15 minutes can be similarly erratic because the contract must adjust to newly opened index stocks and accommodate opening-order flow.

Those session-edge swings distort momentum oscillators and smoothing tools such as moving averages. Session analysis here chooses which minutes remain in the series before the mechanical rules are computed.

How a trimmed window changes a moving average

A 10-period simple moving average computed on a 9:45 to 16:00 window can diverge from the full 9:30 to 16:15 average after the open. The two averages can match again once both lookbacks share the same bars.

Editorial note: the mechanical rules below inherit that same effect, because a changed window means the oscillators and smoothers are no longer the same inputs.

One mechanical day procedure

The archive specified one mechanical trading system: a fully specified entry, exit, and end-of-day flatten tested as one procedure. It combined a 10-period average of a stochastic, 45- and 5-period SlowD smoothing, and a 100-period average of the close. It entered long when that stochastic average crossed above 50 with the close above the long average, entered short on the opposite cross, and flattened at the session close with no stops.

The 50-level cross was specified as trend confirmation rather than 20/80 extreme picking because the design was trend-following. That mid-scale stochastic cross, used with the long moving average of the close, is a trend-confirmation stochastic rather than a fade of oscillator extremes.

What stayed fixed in the evaluation

The evaluation used unadjusted five-minute bars of the current contract. Roll dates were aligned to first-notice dates of the larger related contract. The execution assumptions were a five-dollar round-turn commission and no slippage allowance. Robustness testing then re-ran the same mechanical rules after changing only the session window.

A moderate trim and a tighter window

The same rules were retested after dropping the first and last 15 minutes, then again on a narrower 10:00 to 15:45 window. The tighter window produced weaker evaluation metrics, which the archive treated as evidence that useful information had been removed.

When to trim early

Further reduction that worsens results is presented as a stop signal that the core behavior of the series has changed. Session-window trimming is recommended early in system design only when the noise is consistent.