Over 90 days, analyst targets came clearly closer to the actual price than the model estimates: a typical miss of about 14 percentage points for analysts and 32 for the models. Both were wrong. Analysts were too optimistic in both markets; the models were too high in Finland and too low in the United States. The tests fixed in advance did not settle either way. This is an interim readout: 23 of 64 forecast dates have matured.
This is the measurement fixed in advance. The tables below compare estimates with price moves over 1 or 5 days and are descriptive only. Full result →
Fixed on 25 September 2026, before any outcome existed: is the panel median's mean error closer to zero than the analyst target's, and does a blend with a weight set in advance beat the analyst target alone? Each is checked against the price 90 days later. If the interval includes zero, the result is reported as “no difference established”.
| Model | Directional Accuracy | Mean Absolute Error | Prediction Bias | Hit Rate (±5%) | N |
|---|---|---|---|---|---|
| Claude | 51.5% | 26.8% | +3.9% | 18.7% | 2083 |
| DeepSeek | 51.5% | 30.3% | -0.7% | 12.2% | 2073 |
| GPT | 51.4% | 29.9% | -5.4% | 8.7% | 2094 |
| Grok | 51.1% | 31.2% | -2.6% | 8.4% | 2088 |
| Gemini | 50.8% | 30.4% | +4% | 11% | 1951 |