Drill down — per speaker

Live call — Next FOMC (June 16–17, 2026)

Four central banks, one architecture

Same lexicon engine, same walk-forward backtest protocol, same cross-asset position-sizing. Fed (Powell, n=43) + ECB (Lagarde n=46, Draghi n=41) + BoE (Bailey n=17, Carney n=10) + BoJ (Kuroda+Ueda pooled n=130). The verbal-alpha pattern reproduces on Powell and Lagarde at validated scale. Draghi DAX replicates at +16.0pp on n=41 — second ECB chair, same direction (Sharpe CI still crosses zero, treated as early-evidence). Bailey and Carney show early-evidence edges at small n. BoJ per-chair honestly weak — disclosed in the table below.

Status legend: validated = n ≥ 30 with stable CI · early = n = 10–20, point estimates positive but CIs wide · thin = n < 10, suggestive only · research = pooled signal with weak per-chair components — modest edge survives but we don't claim more than that · weak = per-chair signal honestly does not separate from noise. The FED-built lexicon does not transfer cleanly to BoJ at the per-chair level.

The edge — right tool per asset class

Different asset reactions need different feature sets. Rates pricing is dominated by macro regime × verbal interaction (gradient-boosted on 76 features incl. CPI / NFP / unemployment + verbal). Equity and FX reaction is dominated by verbal cadence alone (linear on 22 verbal features). Honest split, principled, both validated walk-forward.

Edge over momentum baseline (predict same sign as last meeting). Walk-forward, no leakage, n=43 Powell meetings (2018–2026), n=46 Lagarde meetings (2020–2026). Sharpe 95% CI from 5,000-sample bootstrap on per-meeting signal-driven P&L.

Per-speaker signal heterogeneity — pooled fails, individual speakers work

We built a corpus of 520 Fed governor speeches across 2018–2026 (1.47M tokens, 7 governors), and ran the same-day cross-asset model both pooled and per-speaker. The architectural finding: pooling all governors into one signal fails (negative edges across all targets — speakers are not interchangeable). Per-speaker models recover real but modest verbal alpha, with the strongest signal on Waller's rate predictions.

Pearson correlation between Waller's speech-day lexicon score and the realized 5y move = +0.260 on n=63 walk-forward predictions. That's real but modest — what desk research has long suggested about Waller specifically. An earlier 2024-2026-only sample (n=28) showed a stronger +27pp/+0.605 reading; backfilling to the full 8-year history reduced it to the +0.260 above. We report the more conservative number.

The product is per-speaker. Every voting member gets their own corpus + their own model. The desk subscribes to a ranked feed of speaker signals. Pooled-committee aggregation is the wrong architecture and the data confirms it.

vocmarkets — the full verbal-alpha portfolio

The four headline targets above aren't the whole story. Same engine, run against the macro complex: 9 distinct asset classes show positive verbal-alpha edge from Powell's prior-presser features. Credit (LQD/HYG), oil, and BTC join the original rates + equities + FX. Sector rotation (XLF/XLE/XLK) and gold are null — those moves require fundamentals or supply shocks the verbal signal can't see.

Each row is a separate walk-forward backtest, n=43, no leakage. The expanded universe converts a single-asset signal into a portfolio of edges — and the honest null results signal which moves the engine isn't designed to capture.

Direction vs magnitude — two distinct trading-desk products

A direction signal tells you which way to lean. A magnitude prediction tells you how big to size. We measured both honestly. Result: rates predictions are tradeable on magnitude (real MAE improvement over a naive zero baseline). Equity/credit/commodity predictions are tradeable on direction only — magnitudes are correctly-signed but ~2x too aggressive, requiring trader-side shrinkage.

We disclose where the model is well-calibrated and where it isn't. The two products map to different desk subscribers: the rates desk consumes magnitude forecasts; the equity / credit / FX desk consumes direction signals.

Which words actually move equities

For every word in the lexicon, we correlate Powell’s usage frequency at the prior presser with the realized FOMC-day SPX move. The pattern is durable: confident / committed language predicts rallies, hedge / topic-naming language predicts selloffs.

Pearson correlation between word frequency at Powell’s prior presser and the realized FOMC-day S&P 500 move (n=63 meetings).

Call ledger — 3 live, 4 walk-forward backtests

Three live pre-registered calls — BoE Bailey May 8, ECB Lagarde June 5, and Fed Powell June 17 — plus four walk-forward backtest replays of prior pressers. Backtest replays use only data available at the source presser's timestamp — no leakage. They are not pre-registered live calls; they are honest model replays, regenerated from the same `python -m seer_fomc.model.posterior` pipeline. Walk-forward backtest on these 4 entries: 2 hits, 2 misses. The model's broader walk-forward record across all n=43 FOMC meetings is 74.4% argmax-correct (see Calibration section). Bailey resolves in 3 days, Lagarde in 1 month, Powell in 6 weeks.

The June 17 call is committed in this deploy and resolves publicly when the meeting happens. The four backtest entries are walk-forward replays of the model on prior pressers using only data available at the source timestamp (no future leakage); they are not pre-registered live calls. Going forward, every Lagarde / Bailey / Powell call will be pre-registered here before the relevant meeting.

Calibration — when the model says 84% hold, is it actually 84%?

Brier score and log loss measure probability calibration directly. Walk-forward, n=43 FOMC meetings, scored against a trailing base-rate baseline (predict the empirical decision frequency over the prior 12 meetings).

Honest weakness: the model is well-calibrated on hold and hike (the dominant regime in 2018–2024) but under-confident on cuts. It missed the start of the September / October 2025 cut cycle — predicted hold both times when the outcome was a 25bp cut. The cut tail is the part of the distribution that needs more training history, and it's the part of the live June 17 call most worth watching. See every walk-forward prediction →

The product roadmap