The Lab

Every trading day at 9:25 ET the Lab forecasts where SPY will be at 10:00, noon, 2:00 and the close, then learns from what actually happened. The bar it has to clear is the simplest honest forecast: the latest pre-market price holds.

Skill at the close +1.7% ± 1.1% · promising, not proven
Edge hit rate 63% direction beyond the latest price, 63 leaned days
Range coverage 71% target 68%
Days learned 134 first 15 are warmup and not counted

Call

Mon, Sep 28, 2026 Graded 3 inputs missing

Close target $768.04, range $762.55 to $773.56, 31% chance of closing above the prior close ($771.35). Closed at $765.61.

This call was made without everything it normally sees. Each one was read as an average day (no signal), so the call leaned on the rest:
  • Headline risk tone (Jev) missing
  • Event risk today (Jev) missing
  • Fed leaning hawkish (Jev) missing
CheckpointP(up)TargetRangeActualIn rangevs naive
10:00 14% $768.60 $766.72 – $770.48 $767.58 Yes Beat (0.190 vs 0.266)
Noon 25% $768.14 $764.18 – $772.12 $764.95 Yes Beat (0.369 vs 0.429)
2:00 30% $768.31 $763.36 – $773.30 $767.22 Yes Beat (0.226 vs 0.256)
Close 31% $768.04 $762.55 – $773.56 $765.61 Yes Beat (0.267 vs 0.307)
What moved this callWeight times today's reading, for the lead Kalman expert, in sigmas.
pushes downpushes up
  • Gamma above vs below spot
  • Implied vs realized vol
  • Baseline lean
  • Distance to gamma flip
  • Net dealer gamma
  • Room to call wall vs put wall
  • Put skew
  • IWM minus QQQ yesterday

Track record

Is it learning? Skill vs "latest price holds"Cumulative CRPS skill at the close, with a 2-standard-error band. Above zero means better than naive.
-10.0%-5.0%+0.0%+5.0%+10.0%
naive+1.7%
May 22Jul 20Sep 11Nov 4
Is the range honest?Share of the last 20 closes inside the displayed range. The Lab steers this toward 68%.
40%60%80%100%
68% target75%
May 22Jul 20Sep 11Nov 4
Who the Lab trustsEnsemble weight per expert at the close. Experts earn weight by forecasting better than the rest.
  • Kalman: everything, slow memory 16%
  • Gaps keep going 13%
  • Analog days (10 nearest) 11%
  • Kalman: price action 10%
  • Kalman: options positioning 10%
  • Everyone else 40%
0%50%100%
Apr 27Jun 30Sep 2Nov 4
CalibrationWhen the Lab said X% chance of an up close, how often it was
0%50%100%
24
16
16
19
44
0%25%50%75%100%

Leaderboard

Average CRPS at the close in sigma units (lower is better), and each expert's edge over "latest price holds". Weight is how much the ensemble listens to it today.

ExpertCRPSvs naiveWeightVerdict
Kalman: everything, slow memory 0.4693 +0.0094 ± 0.0099 16.5% ~408 more days to confirm
Gaps keep going 0.4714 +0.0073 ± 0.0078 12.7% ~433 more days to confirm
Kalman: options positioning 0.4748 +0.0039 ± 0.0076 9.6% ~1686 more days to confirm
Kalman: price action 0.4763 +0.0024 ± 0.0096 9.6% ~7708 more days to confirm
Analog days (10 nearest) 0.4770 +0.0017 ± 0.0078 11.2% ~10103 more days to confirm
Latest price holds (random walk) 0.4787 — 7.8% the benchmark
Yesterday's move reverses 0.4803 -0.0016 ± 0.0047 6.0% no edge so far
Kalman: Jev text reads 0.4810 -0.0024 ± 0.0049 7.1% no edge so far
Yesterday's direction continues 0.4839 -0.0053 ± 0.0048 7.2% no edge so far
Dealer hedging drift 0.4844 -0.0057 ± 0.0045 5.5% no edge so far
Kalman: everything, fast memory 0.4869 -0.0082 ± 0.0148 4.6% no edge so far
Gaps fade 0.5028 -0.0241 ± 0.0085 2.2% no edge so far

By checkpoint

CheckpointSkillFrom rangeFrom edgeCoverageEdge hit rate
10:00 +9.4% ± 2.2% +2.1% +7.2% 72% 75% (73)
12:00 +1.5% ± 1.0% +0.7% +0.8% 69% 68% (40)
14:00 +2.4% ± 0.9% +2.0% +0.4% 71% 54% (35)
Close +1.7% ± 1.1% +0.7% +1.0% 71% 63% (63)

How it learns

  • Senses. Options positioning from the 9:25 snapshot, yesterday's moves in stocks, bonds, credit, the dollar and VIX, and Jev's reads of the morning headlines and WSB.
  • Experts. Simple fixed hypotheses ("gaps fade"), Kalman filters whose weights drift slowly with the market, and an analog expert that asks what happened on the most similar past days.
  • Ensemble. Experts earn weight by forecasting better (Hedge), and every expert keeps a small share so one that fell out of favor can come back when the market changes.
  • Range. The range widens after a close lands outside it and narrows after one lands inside, so over time about 68% of closes fall inside it.