Why a Stock Alert Win Rate Above 50% Can Still Have Negative Expectancy

A win rate answers a narrow question: among evaluated outcomes classified as wins or losses, how often did a win occur? It does not tell you how large the wins and losses were—or what a subscriber earned.

Foreshadow AI’s own evaluated-alert snapshot illustrates the distinction. Looking at an unfavorable result alongside its methodology is more useful than presenting a favorable percentage alone.

A real example from our evaluated alerts

The website worker snapshot generated on September 16, 2026, at 20:21:42 UTC reports the following for hourly alerts in its rolling 365-day lookback:

| Metric | Recorded value, rounded | | — | —: | | Evaluated alert count | 160 | | Win rate, excluding pushes | 52.5% | | Average winning move | 0.8671% | | Average losing move, magnitude | 1.0128% | | Expectancy per evaluated non-push alert | −0.0259 percentage points |

These are historical evaluated alert outcomes, not executed subscriber returns. The window is selected by evaluation check timestamp. It is not a completed calendar year or a record of trades executed during that year.

The important combination is a win rate above half and negative expectancy. Both can be true because the average losing move was larger than the average winning move.

Follow the arithmetic, not just the headline

For these non-push outcomes, the relationship is:

Expectancy = win fraction × average winning move − loss fraction × average losing move.

Using the displayed, rounded figures:

0.525 × 0.8671 − 0.475 × 1.0128 ≈ −0.0259 percentage points per evaluated non-push alert.

This is an explanation of the recorded metric, not a forecast. It does not mean an account lost 0.0259% each time an alert appeared. The calculation does not establish subscriber execution, position sizing, overlapping exposure or portfolio performance.

A win rate without the sizes of wins and losses leaves out information needed to interpret the result.

Check what the calculation leaves out

This evaluated-alert metric does not deduct commission or slippage. Its hypothetical gross signal outcomes are not independently verified execution or subscriber returns, and the snapshot is not broker verified.

The snapshot also reports mixed methodology versions across its evaluated-alert history: 2,152 version-1 rows and 107 version-2 rows. Those counts describe the reported history, not the hourly segment alone. A segment-specific version breakdown is not supplied.

Version 2 uses causal completed-session bars, the first bar open at or after the alert, a gap-aware hypothetical stop, the full holding period and close timestamps. Legacy version-1 rows remain and have not been restated. Mixed-version history should not be described as one uniform strategy.

These limitations matter even when a headline number looks attractive.

Copy this evidence checklist

Before interpreting a performance figure, record:

Unit: Evaluated alerts, recorded closed trades or verified account returns?
Window: A rolling lookback or a completed calendar period? Which timestamp selects records?
Sample: How many observations, and what does the win-rate denominator exclude?
Magnitude: What are the average winning and losing moves?
Costs: Which costs are deducted, and which remain unverified?
Methodology: Did the evaluation rules change within the history?
Boundary: What does this evidence not establish?

Keep different trackers separate. Do not combine these evaluated-alert figures with a closed-trade tracker or convert them into a subscriber-return claim. Also, the evaluated-alert snapshot defines “average move” as absolute movement—not signed profit or account growth.

Evaluate research fit separately

Foreshadow AI provides research and alerts covering 520+ U.S. equities, with delivery through Telegram and Discord. We do not execute subscriber trades. Confidence scores are not calibrated probabilities of profit; they are not interchangeable with the historical win rate discussed here.

Use our seven-day trial to assess whether the research fits your review routine and whether you can explain its limitations. You do not need to place trades to evaluate those questions. A short trial cannot establish dependable investment-performance expectations.

Compare plans and start a seven-day trial. Review the terms presented before subscribing.

References

Foreshadow evaluated-alert performance — figures and methodology from the dated worker snapshot above; subsequent page values may change.
Five questions for interpreting confidence scores — a companion research checklist.

Foreshadow AI provides market analytics and alerts for information and education. It does not execute trades for subscribers. Investing involves risk, including loss of capital. Past performance does not predict future results. Confidence scores are not calibrated probabilities of profit.

Explore Foreshadow plans and start your seven-day free trial

Why a Stock Alert Win Rate Above 50% Can Still Have Negative Expectancy