Cleaner Signals Can Still Be Worse Signals

A model that produces fewer bad alerts may look improved—until we count how many valid events it stopped finding.

“Cleaner” is an appealing word in automated trading research. Fewer signals can make a chart easier to read, reduce turnover, and lower the number of obvious mistakes. But fewer events are not automatically better events.

Our causal early-turn pivot experiments repeatedly exposed this tradeoff. Several designs reduced false events, yet they did so largely by suppressing real peaks and troughs. Because our computer-vision research log is append-only, those rejected versions remain in the record rather than disappearing behind the latest model.

Evidence classification: The results below include fresh out-of-sample detector diagnostics, one sealed blind out-of-sample comparison, and chronological development tests. They do not establish trading profitability.

The first early-turn model looked cleaner

Our pending-turn v1 detector was designed to remove the four-bar availability delay of an offline radius-4 benchmark. On a fresh July manually labeled window, its false-event rate improved from 6.63 to 3.32 events per day.

Viewed alone, that reduction looks encouraging. The broader scorecard told a different story:

• Zone-hit rate fell from 74.19% to 65.22%.
• Peak F1 fell from 48.70% to 17.81%.
• Trough F1 fell from 51.69% to 23.13%.

V1 emitted fewer false events because it emitted too few valid events as well. The visual output was cleaner, but the detector was less useful.

Earlier signals can fail in the opposite direction

We then froze a segment-safe v2 model and evaluated it once on a newly labeled blind window. V2 reached events with a two-bar median lead relative to radius-4 availability. It also produced 1.79 times as many tolerance false events as the benchmark.

That model failed its frozen preservation gate and was rejected. The earlier timing did not compensate for the increase in false events. Because the window had now been scored, it became consumed development evidence; we did not retune v2 and call the same window blind again.

Together, v1 and v2 showed both sides of the problem. One design became too quiet and missed valid turns. The other preserved an earlier lead but became too noisy.

More filtering repeated the low-recall failure

Two later development experiments tested narrower hypotheses under separate precommitments. V3 changed probability re-arming. V4 used a binary event gate followed by side classification.

Both substantially reduced false events on chronological development folds. Both also repeated the low-recall failure mode. Neither was frozen for a new blind package, and downstream trading tests remained closed.

This is a useful negative result. If successive filters improve chart cleanliness mainly by withholding detections, adding another filter is unlikely to solve the underlying observability problem.

A useful scorecard must show what disappeared

We now resist summarizing detector quality with signal count or false-event rate alone. At minimum, the same held-out rows should report:

• event count and false events per day;
• precision, recall, and F1 for peaks and troughs;
• zone hits and strict one-to-one matches;
• side confusion and lead/lag relative to the benchmark;
• the exact time each candidate became causally available.

The scorecard should also retain separate “candidate,” “available,” “signal,” and “entry” timestamps. Otherwise, a clean detector chart can conceal an unusable execution delay.

Clean is an outcome, not an objective

The goal is not to minimize alerts. A detector that emits nothing has zero false events and zero value. The goal is to preserve enough real information at a live-safe time for the next research layer to have a chance.

Our early-turn tests have not yet cleared that gate. That conclusion is more useful than selecting the prettiest chart, because it tells us where not to spend the next round of strategy work.

Follow the next research updates on the Evolve Quant blog or join our newsletter.

Research status as of August 5, 2026. Later evidence may narrow or supersede this interpretation.

Important: This article is for educational and research purposes only. It is not investment advice, a trading signal, or a promise of future performance.