Market bias
A score from −10 to +10 per asset and horizon, computed continuously from the news, the market regime and quantitative models, with its measured accuracy.
Measured accuracy
No better than chance so far: the confidence interval contains 50%.
Naive baselines on the same calls: always long 51.5%, previous direction 51.8%.
Version 1, published until Sep 25, 2026 (live tracking, Jul 20 → Sep 25, 2026): 49.6% over 1,698 calls, CI 47% to 52%.
Breakdown by asset and horizon
| Asset | 0 – 30 min | 1h – 4h | 1d – 4d |
|---|---|---|---|
| NQ / MNQ | 54.4%371 calls · CI 49% to 59% | 57.1%105 calls · CI 48% to 66% | 44.7%38 calls · CI 30% to 60% |
| ES / MES | 54.1%333 calls · CI 49% to 59% | 50.5%105 calls · CI 41% to 60% | 56.4%39 calls · CI 41% to 71% |
| YM / MYM | 52.9%291 calls · CI 47% to 59% | 44.1%93 calls · CI 34% to 54% | 55.6%36 calls · CI 40% to 71% |
| GC / MGC | 50.6%239 calls · CI 44% to 57% | 59.3%54 calls · CI 46% to 71% | 57.1%35 calls · CI 41% to 72% |
| CL / MCL | 57.1%14 calls · CI 33% to 79% | 65.0%40 calls · CI 50% to 78% | 62.5%32 calls · CI 45% to 77% |
| NASDAQ | 55.3%161 calls · CI 48% to 63% | 82.6%23 calls · CI 63% to 93% | 55.0%40 calls · CI 40% to 69% |
| DOW JONES | 51.5%134 calls · CI 43% to 60% | 61.9%21 calls · CI 41% to 79% | 59.0%39 calls · CI 43% to 73% |
| S&P 500 | 55.1%156 calls · CI 47% to 63% | 77.3%22 calls · CI 57% to 90% | 56.8%37 calls · CI 41% to 71% |
| CAC 40 | 45.9%109 calls · CI 37% to 55% | 55.6%27 calls · CI 37% to 72% | 61.1%36 calls · CI 45% to 75% |
| DAX | 52.7%131 calls · CI 44% to 61% | 57.1%42 calls · CI 42% to 71% | 53.8%39 calls · CI 39% to 68% |
Data as of Sep 25, 2026, 21:04 (Paris time)
Method
Not predictive if the hit rate is below 52% or if the 95% confidence interval contains 50%.
Non-overlapping windows: 0 – 30 min every 30 min (xx:00 and xx:30 UTC) · 1h – 4h every 4 hours (00, 04, 08, 12, 16 and 20 UTC) · 1d – 4d one session (14:00 UTC to the next weekday 14:00 UTC). A call = a score of at least 0.8 in absolute value at the start of the window; it is right if the price ends the window in its direction.
Naive baselines: “always long” and “same direction as the previous window”, on the same windows.
Historical replay: scores recomputed after the fact on the headlines recorded at the time. The v2 rules were written while reading headlines from this same period: this is not an out-of-sample result.
- Only the headlines that version 1 rated important enough were archived: a few slightly less important headlines, which version 2 would keep, are missing from the replay.
- The number of outlets that picked up each headline was not archived: in the replay, every headline counts as if it had been published only once.
- The language model's opinion on each headline was only kept from September 21 to 25: the main replay therefore uses the lexicon alone. The variant in which this model can set a headline aside is only studied over those five days, and only for the distribution of scores.
- The version 2 lexicon rules were written while reading headlines from this same period: the replay is therefore measured on the data used to build it, not validated on new data.
- The 10 assets often move together (NQ and the Nasdaq, ES and the S&P 500, YM and the Dow follow the same market): calls added up across assets are not independent.
- The quant and regime pillars are taken as they were recorded at the time, not recomputed.
- Yahoo prices with the regulatory delay (about 10 minutes on CME futures), as displayed; continuous contracts, without correcting the jumps at each contract roll.