Event-predictability versus position-constrained rules

Two exams of the same rule set

A statistical design workflow first builds a dataset of defined events, then tests whether subsequent price change over a fixed bar count has a directional tendency. That measurement is event-predictability: the tendency of a computer-defined occurrence to be followed by a price change over a fixed number of bars.

TradersWeek editorial reading: treat system evaluation as two exams of the same rule set. Rank every defined event first on its unoccupied lookahead distribution. Then force those controls through a position-aware engine and a later or opposite-regime sample, so the occupancy-filter, stops, and walk-forward-analysis can overturn the first ranking.

A close-based stochastic as the defined event

The demonstration event is a close-based stochastic, formed from the latest close relative to the highest and lowest closes in a lookback window.

Several reconstructions enter long on a stochastic threshold cross and exit either after a fixed holding period or when price hits a percentage stop from the entry. Those exits are a timed-exit and a percent-stop.

Shared default controls in several implementations are a lookback of 8, a threshold of 0.3, a 14-bar hold, and a 3.8 percent stop.

How event-predictability is scored

Event scoring looks ahead a set number of bars after each occurrence, bins the percent change, and summarizes the distribution with a center-of-gravity weighted average. A more positive reading is treated as more favorable for long-side use.

A scenario tester can sweep stochastic lookbacks from 8 to 18, thresholds from 0.1 to 0.35 in 0.05 steps, and lookaheads from 5 to 18, then rank combinations by center-of-gravity.

When occupancy and stops change the ranking

Many events are ignored while a position is already open. That occupancy-filter means the engine does not take every occurrence that entered the event-predictability sample.

Stop-loss processing can prevent a taken trade from realizing the lookahead path used in the event score. The percent-stop can close the trade before the fixed bar count that produced the center-of-gravity reading.

System-optimization is a joint search over entry, exit, and abstention controls. The full procedure, not a single indicator setting, is what gets scored.

A later or opposite-regime retest

One reconstruction warns that a 10-year in-sample optimization window was a strong bull market, so robustness should be rechecked on a different subset that includes a bear market.

Walk-forward-analysis is that later or differently regime-shaped retest of controls chosen on an earlier window. The in-sample-regime used to choose controls may not represent later or opposite market conditions.

The event definition can be swapped

The distribution test can swap the event definition for another rule. A larger higher-frequency sample produces a smoother, more bell-shaped histogram than a smaller sample.