A combined forecast weighting the AI panel at 25% and analysts at 75% was less wrong than either alone: mean absolute error 12.6% vs analysts 14.1% and AI 24.6%.
Out-of-sample check: blend weight fitted on data up to 2026-05-05, then evaluated on the 575 later observations — blend 12.6% vs analysts 14.3%.
Error as a function of the blend weight
Each point = mean absolute error when the forecast mixes AI and analysts at that weight. The dip between the endpoints is the whole story: the blend beats both.
optimal weight (in-sample)out-of-sample check
AI panel bias
-8.5%
Mean forecast error of the AI panel median (no analyst blending) vs the realized price. Negative = systematically conservative.
Analyst bias
+8.8%
Mean forecast error of the analyst consensus target. Positive = systematically optimistic — consistent with the documented analyst-optimism effect.
Scorecard · n=1401 · 2026-03-17–2026-06-09
AI panel
Analysts
Blend (25% AI)
Mean absolute error
24.6%
14.1%
12.6%
Directional accuracy
52.2%
56.6%
60.2%
Systematic bias
-8.5%
+8.8%
—
Engine v8 — separate window · n=1642 · 2026-06-10–2026-09-16
AI panel
Analysts
Mean absolute error
30.4%
12.7%
Directional accuracy
52.5%
65.7%
Systematic bias
-5.7%
+10.1%
Engine v8+v8.1 forecasts, scored separately: the v8 cap reform changed the estimate distribution, so these windows are not comparable with the v6+v7 scorecard above. Early sample — treat as indicative.
Forecast error over time
Daily mean forecast error (estimate vs realized price at the horizon). 0% = unbiased.
AI panelAnalysts
By market regime
n
AI bias
Analyst bias
AI MAE
Analyst MAE
Up-market windows
711
-9.7%
+7.0%
25.3%
13.9%
Down-market windows
690
-7.2%
+10.7%
24.0%
14.3%
Regime = sign of the median realized return across tracked companies in each forecast window (derived from our own data, no external index).
Read this before quoting the numbers
The AI series is the pure-model panel median — the models' own view with no analyst blending. Since v8.1 (2026-07-14) this is the same series as the site's headline consensus; before that the headline was a calibrated blend (scored separately as a research series).
Main measurement window: engine v6+v7. Engine v8-family forecasts are scored in their own block above (different cap regime, not comparable). 528 v8 forecasts are still maturing — the next windows mature 2026-10-17.
The optimal blend weight is fitted in-sample; the out-of-sample line above is the honest check. Treat both as indicative — the sample is a few months, not years.
Analyst targets come from Yahoo Finance and update with a lag, which flatters analyst responsiveness slightly.
Forecasts are 12-month targets scored at a 30-day horizon — this measures direction and calibration, not whether the 12-month target ends up exact.