AI · Berkshire Mirror — Track record
Every elapsed decision is graded against what prices actually did — hit-rates and errors per signal family, plus the cost of the decisions this book chose not to take.
53 graded 1 pending
1W calibration 0.03 How far along the 1-week path a decision typically gets: the fraction of the decided move realized after one week, measured across this book's graded decisions. Feeds a bounded conviction adjustment back into future decisions.
Scorecard
Hit-rates per signal family.
| Signal family | Dir | n | Hit 1W | Hit 4W | Hit 12W | Med err 1W | Stop-first | Bias | Status |
|---|---|---|---|---|---|---|---|---|---|
| FailedBreakout_Down_20D | sell | 22 | 53% | 33% | — | — | 0% | +1 | ● active |
| NoTrade | sell | 10 | 25% | — | — | — | 0% | -5 | ● active |
| PullbackBuy_VWAP | buy | 6 | 50% | — | — | -1.7% | 50% | 0 | ● active |
| RallySell_BearTrend | sell | 6 | 17% | — | — | — | 0% | -7 | ● active |
| BreakoutUp_20D | sell | 3 | 67% | — | — | — | 0% | 0 | ○ building sample (3/5) |
| BreakdownDown_52W | sell | 2 | 0% | — | — | — | 0% | 0 | ○ building sample (2/5) |
| MomentumContinuation_Up | buy | 2 | 0% | — | — | -3.9% | 0% | 0 | ○ building sample (2/5) |
| FailedBreakdown_Up_20D | sell | 1 | — | — | — | — | 0% | 0 | ○ building sample (1/5) |
| FailedBreakdown_Up_2 | sell | 1 | 0% | — | — | — | 0% | 0 | ○ building sample (1/5) |
Counterfactuals Counterfactuals grade decisions that were NOT executed — expired queues, no-chase rules, untaken optional trims — against what the market then did. They are kept strictly separate from the calibration cells above.
What the decisions we didn't take would have done.
- Optional trims not taken — 16 cases , median would-have 1W move -0.6% → verdict: cost
- No-chase queue (queued a pullback instead of chasing) — 8 cases , median would-have 1W move +3.0% → verdict: cost
Counterfactuals are advisory evidence for the book's learnings reviews — never blended into the calibration cells.