Keep three jobs separate when you test a composite score

A compressed score is still three jobs

A proprietary-model compressed more than 30 technical indicators into one composite-score scaled from 0, labeled very bearish, to 100, labeled very bullish. Score histories for 10 securities covering May 30 through October 18 were supplied together with a 10-day exponential moving-average of those scores.

A moving-average is a lookback smoother of ordered observations that damps short swings in a score or price series so the series can be compared with later outcomes. A trend-filter is a classification of whether a series is rising, holding a range, or falling, used to read continuation versus change instead of issuing a single flip. A mechanical-trading-system is a fully specified entry, exit, and abstention procedure that turns a score or filter state into a testable signal over a defined holding period.

Editorial: the composite-score is the compressed input. The smoother, the trend reading, and the trigger remain separate procedures, even when a service offers them as one number.

What happens when the score is the trigger

Using the composite-score itself as a mechanical buy and sell trigger produced late sell indications in the reviewed histories. In one illustrated name, the score stayed bullish through a consolidation and into a decline, with a sell indication arriving only after price had already moved down to 122.

Reading rising scores as bullish, range-bound scores as continuation, and falling scores as bearish treated the same series as a trend-filter and produced mixed results that the reviewer judged better than binary score triggers. In another illustrated name, a moving-average of scores that remained in a high range coincided with continued strength from May 30, 2000.

Editorial: a late sell after a still-bullish consolidation is evidence about how the mechanical-trading-system was specified, not evidence that the market failed the score.

A proprietary model cannot finish the audit

Because the scoring and screening rules were proprietary, the composite could not be independently reconstructed or audited. The service was positioned as a way to avoid doing technical analysis, yet the reviewer argued that using the scores still required that knowledge.

Editorial: if the internals of a proprietary-model cannot be checked, write the moving-average, the trend-filter, and the mechanical-trading-system as three testable jobs. Then a late exit can be blamed on test design rather than on the market.