Why Pivot Detection Is Not a Trading Signal

Finding a recognizable peak or trough on a chart is a detection problem. Turning it into a live-safe, profitable decision is a much longer chain.

Our ETHUSD pivot research began with a simple question: can a model locate the kinds of turns a human annotator marks on a five-minute chart? The answer is partly yes. But our results also show why pivot detection should not be described as a trading signal.

A detector can score well around a visual turning zone and still arrive too late, emit too many false events, rank the wrong events, or lose money after entries, exits, latency, and trading costs. Those are separate tests, and we record them separately in our append-only computer-vision research history.

Evidence classification: The metrics below are fresh manually labeled out-of-sample detector diagnostics and causal timing diagnostics. They are not P&L evidence.

Our offline benchmark recognizes zones better than exact bars

On a fresh July evaluation window, our radius-4 benchmark achieved a 74.19% zone-hit rate. Its exact F1 score was 21.56%.

Those two numbers answer different questions. A zone hit gives credit when a predicted event lands close to a manually labeled turn. Exact F1 requires more precise bar-level agreement while balancing precision and recall. A chart can therefore look directionally sensible even when exact timing remains difficult.

Neither metric says whether the event could have been known at the displayed timestamp. It also says nothing about trade direction, entry rules, stop placement, holding period, position sizing, or net return.

Visual event time is not information-availability time

The radius-4 benchmark uses four future five-minute bars to shape an event. That makes it useful for offline visual localization, but not live-safe at the pivot timestamp. The system cannot know those future bars until roughly 20 minutes have passed.

When we traced that event through a causal confirmation and entry lifecycle, the shaping delay consumed approximately 0.3681% to 0.3907% of price movement before executable entry. This was a timing diagnostic, not a universal slippage estimate, but it exposed the central problem: a visually accurate label can describe a turn that is no longer available to trade at the labeled price.

This is why every event needs multiple timestamps: where the turn is assigned, when the candidate becomes knowable, when a signal is emitted, and when an order could realistically be executed.

Detection is only the first layer

We separate the research into distinct evidence categories:

1. Detector accuracy: Did emitted events match manually labeled peaks and troughs?
2. Event ranking: Did higher scores identify better candidates among emitted events?
3. Causal timing: Were all inputs available when the decision was made?
4. Lifecycle behavior: How were candidates updated, confirmed, canceled, entered, and exited?
5. P&L: Did the complete strategy produce positive results after realistic costs and risk controls?

A result in one category cannot substitute for the next. High recall does not establish good event ranking. Good ranking does not repair future leakage. A causal signal does not guarantee a workable entry. A promising gross return does not establish a durable net edge.

A live-safe detector faces a different tradeoff

We tested causal early-turn alternatives designed to act before the radius-4 benchmark became available. Some arrived earlier, but they gave up too much accuracy or created too many false events. That is not surprising: earlier decisions have less evidence.

The right comparison is not “early” versus “late” in isolation. It is whether a frozen early detector preserves enough held-out precision and recall, controls false events, maintains sensible side classification, and improves executable timing on the same rows. Only then should downstream lifecycle and trading research reopen.

What would make a pivot actionable?

An actionable system would need to pass held-out event matching, prove causal availability, define a robust stateful lifecycle, and survive chronological tests with fees, spread, slippage, latency, overlap, exposure limits, and failure controls. It would also need enough independent events to show that the result is not carried by one short market regime.

Until those layers pass, our pivot work remains detector research. The most accurate statement is not “AI predicts ETH turns.” It is that we have an offline benchmark with measurable localization ability and an unresolved live-timing problem.

Read more development findings on the Evolve Quant blog or join our newsletter for future research updates.

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.