Methodology

Predict the probability distribution of the US headline CPI YoY print (NSA, rounded to 0.1pp — the number prediction markets resolve on) better than the market prices it. The market price is the consensus; matching it is worth exactly nothing. Everything on this site is scored against that bar.

Kill criterion: if the model's log-loss on realized prints does not beat the market's over a trailing nine live cycles at T−7, there is no edge and the project should be killed or rethought. The scoreboard tracks this continuously.

The point estimate is the Cleveland Fed daily inflation nowcast. The model's contribution is honesty about that estimate's error as a function of time-to-release, fitted from ~150 resolved months of daily nowcast vintages back to 2013:

h      = days until BLS release
μ      = nowcast(today) − bias(h)     # trailing 24-month mean error
σ      = rmse(h)                      # trailing 24-month RMSE, floor 0.08pp
df, s  = Student-t MLE fit on standardized vintage errors (point-in-time)
p_t    = ∫ t(df, μ, s·σ) over bucket, normalized
p_pub  = λ·p_t + (1−λ)·p_market       # λ fit on resolved prints, [0.2, 0.8]

There is no regression, no hand-tuned constant, and no conviction heuristic. The uncertainty claimed at T−12 is the error the nowcast actually made at T−12 over the trailing window; the window tracks the volatility regime (2021–23 errors were ~2.5× the 2013–19 ones), and the floor is the best T−7 RMSE the nowcast ever achieved in its calmest regime — we never claim more precision than the source has ever had. A useful empirical fact from the fit: the nowcast's error barely improves inside the final three weeks (RMSE ~0.10–0.11 from T−21 to T−1), so the model's information advantage, if any, does not grow by waiting.

v2 (after the June 2026 print, a +0.39pp nowcast miss during the Iran-war energy break): the error distribution is a Student-t, not a normal — standardized vintage errors show ~4% of months beyond 3σ vs the Gaussian's 0.27% — and the published distribution blends the model with the market's own prices at a weight λ fit point-in-time on resolved prints. Blending bounds the damage of any single model blow-up at ln(1/(1−λ)) relative to the market while preserving the model's information where it has any. An intramonth-gasoline correction was tested for the same release and rejected: over 154 months the nowcast's error is essentially uncorrelated with the month's gasoline move (r ≈ −0.15) — the Cleveland Fed model normally absorbs energy; June was an idiosyncratic failure, which is precisely what fat tails and the blend are for.

An earlier version (l1-ridge-v0.1) fit a ridge regression on top of the nowcast; its out-of-sample skill was negative and it was removed. If a future feature (component decomposition, energy pass-through, trailing-rent leads) can't beat hcal on the scoreboard, it doesn't ship either.

Every resolved market is scored at fixed horizons (T−7, T−2, T−1): the model's probability on the bucket that actually printed vs the market's normalized price on that bucket at the same moment, as log-loss and Brier score. Retro months are reconstructed point-in-time: only the nowcast vintage visible that day, and a horizon curve fitted only on months already resolved by then.

Log-loss against the market accrues evidence every print. A trade-level record (win/lose on monthly binary events) needs years to distinguish skill from luck — twelve prints a year, most correctly untraded. Distribution scoring is the primary metric; paper P&L is color.

There is no approval step and no conviction gate. Daily, for each active market, the paper trader computes the fee-adjusted edge per bucket and either enters or explicitly declines — and logs the full judgment (inputs, skipped buckets with reasons, sizing) either way. The dashboard shows today's decision; the record is override-after-the-fact, not ask-before-acting.

admit bucket if:
  odds fresh (≤2d) and volume ≥ $1k
  entry price (ask, fallback mid) in [2¢, 95¢]
  edge = p_model × 0.98 − entry ≥ 5pp
size: proportional half-Kelly, total ≤ 12.5% of bankroll
window: T−14 to T−3 (inside T−3 the market is sharpest)

The 2¢ floor exists because sub-2¢ buckets are lottery tickets whose spread and fee eat the theoretical edge, and a record built on them is one lucky 30:1 fill wearing a track record. Positions are paper only, held to resolution, settled at 98¢ per winning share (Polymarket's 2% fee on winnings).

Before each release the pipeline writes a one-page brief: the call, the full model-vs-market distribution, what each bucket implies for the month-over-month print given base effects, the nowcast trajectory, the scoreboard prior, the decision taken, and the freshness of every input. The point is an evidence chain — any headline number on this site can be traced to the raw series that produced it, and disagreements with the market are stated as falsifiable real-economy claims before the print, not after.

• The realized-YoY reconciliation uses current revised BLS values, which can differ from the first print by a rounding step; markets resolve on the first print. First-print vintages (ALFRED) are a known future fix.
• The Student-t tail fit and blend weight are estimated on ~150 months of vintages and ~20 odds-covered prints respectively; the blend weight in particular is a noisy estimate and is bounded to [0.2, 0.8] for that reason.
• Historical odds are daily Dune snapshots (mid only); backtest entries assume fills at mid, which flatters the backtest by roughly the spread.
• The scoreboard's retro months include wide companion markets (e.g. "≥2.8%") where both sides score near-perfectly; they dilute the comparison symmetrically but shrink the apparent gap.
• Rent indices (Zillow, ApartmentList) and the Manheim series are ingested but currently unused by the model; they are candidate features gated on beating hcal on the scoreboard.