Warning Scoreboard
We don't predict prices. We flag overheating — every day we log the pool's strong crowding signals, then verify against actual prices 30 days later, fully public. The follow-through rate is this product's only advertising.
🔬 Engine self-check
Every call is baselined the moment it is made and judged automatically after 3 trading sessions · market-adjusted (excess vs SPY)
The interval spans 50% — at this sample size it is statistically indistinguishable from a coin flip. Not yet evidence of skill. About 400 judged samples are needed to resolve a 5-point difference.
Magnitude calibration
10%
Realized ratio · of the claimed move actually travelled
5.5
Mean absolute error · percentage points
55%
Overclaim rate · share that fell short of half the claim
The engine systematically overstates magnitude: the median move reaches only 10% of what it claimed. This measure is more robust than hit rate (continuous rather than binary) and maps directly to what we promise — how big the impact is.
Post-hoc verification of engine calls, not investment performance · not investment advice
Overheat warnings (total)
2026
Expired & verified
311
Methodology
Crowding ≥60 logs an overheat warning; 30 days later we check whether price moved against the crowded side (crowded-long → down / crowded-short → up)
Why this yardstickBacktested on two non-overlapping windows (Feb–Jul 2026)
Out-of-sample · strong-call reversal (T+21)
72.5%
n=69 · excess return vs SPY +6.47%
In-sample · strong-call reversal (T+10)
74.0%
n=50 · excess return vs SPY +4.46%
Calls that news flow makes look 'obvious' systematically reverse within 2–4 weeks — mean reversion after attention-driven overreaction. Replicated in both independent windows, stronger at longer horizons. That's the statistical basis of the crowding warning: not that we can predict direction, but that overheating itself has a measurable tendency to reverse.
Limitations: both windows sit in 2026 (a single regime); observations cluster by news, so the effective sample is smaller than nominal; the numbers above are research results. The live verification on this page starts from scratch with the log above — it does not inherit the backtest.
Evidence track recordEvery piece of evidence the conviction engine cites, independently verified after 30 days (excess vs SPY)
| Evidence type | n | Direction hit rate | Avg excess |
|---|---|---|---|
| 📰 News direction | 930 | 51.6% | +0.2% |
| 🔥 Crowding | 506 | 51% | +0.63% |
| flow | 483 | 48.9% | -0.78% |
| 🧑💼 Insider | 448 | 51.6% | -0.09% |
| 🏗️ Structural | 291 | 49.8% | +1.1% |
| 📊 Market-implied | 237 | 57.8% | +0.6% |
| 💰 Priced-in | 75 | 62.7% | +1.01% |
Overall stance record (bullish/bearish convictions): 127/253 (50.2%) · -0.48%
Methodology: every day at 00:00 UTC we snapshot crowding for every stock in the pool; scores ≥40 are logged immutably; overheat warnings (≥60) are verified against actual closing prices at 30-day expiry. MVP uses absolute prices; the SPY-excess yardstick ships with the production version. For investment research only; not investment advice.