TL;DR: We have tested whether our US CPI forecast beat Bloomberg Consensus and Kalshi. The result is clear: Our forecasts contain information that is not fully reflected in either Bloomberg Consensus or Kalshi prices. Our bottom-up is the most accurate forecast 10 days before the CPI release, while disagreements between our proprietary models and Kalshi provide a directional signal that can be translated into risk-controlled trades (some signals have 80%-100% win rate – see below). We have also developed a model to forecast US gasoline prices which can be used to trade on Kalshi.
(We have a new website – www.underlyinginflation.com. All models, including our forecasts, are now available online, always up-to-date. The interactive dashboard includes a tool to build your own scenario. We will announce it in the coming days, you are welcome to take a look)
WHAT WE DID
Our exercise compares two independent proprietary forecast signals against markets expectations. The first is our bottom-up US inflation forecast, constructed from the underlying components of the CPI basket. (Just to be clear: this is the US CPI forecast we circulate around 10 days before the CPI release.) The second is our Machine Learning (ML) model, estimated recursively so that every historical prediction uses only the information that would have been available at the time. We evaluate both against Bloomberg Consensus forecasts and against Kalshi prices. Consensus is the natural survey benchmark; Kalshi is the more demanding trading benchmark because its prices update continuously and directly determine the cost of a position. The analysis covers headline and core CPI, both month on month and year on year. All data are real-time data. The common four-model sample begins in January 2023 and ends in June 2026, producing 147 matched observations. Every comparison is made on the identical panel.
We examine snapshots from 30 days before the release through 14 days, 10 days, one day and one hour. (For Consensus we use only the 1h-before data.) The performance of the forecast is assessed with mean absolute error, the dispersion and maximum size of the forecast miss, and the frequency with which the forecast matches the published number at one decimal. Not only, but we then ask a more conditional question: when Consensus or Kalshi is ultimately surprised, did the proprietary model identify the direction of that surprise? Finally, we convert the signals into executable YES and NO Kalshi positions using recorded bids, asks, taker fees and feasible exit horizons.
RESULTS
MEAN ABSOLUTE ERROR
Ten days before release, bottom-up has the lowest pooled mean absolute error, at 0.070 percentage point, versus 0.077 for Consensus, 0.082 for ML and 0.097 for Kalshi. It also has the highest exact one-decimal rate and the smallest maximum miss. Both MoM and YoY of our models are competitive for core and headline CPI. Headline forecasts of our models are compettive: bottom-up records a MoM MAE of 0.069 and a YoY MAE of 0.070, while the YoY MAE of the ML model (0.084) is below Consensus (0.090) and Kalshi at (0.092). As for core, the YoY MEA of bottom-up (0.71) is below Consensus (0.74) and Kalshi at any date/horizon. More broadly, both proprietary models retain an advantage over Kalshi in headline and core year-on-year inflation even one hour before the release.
(Please, note that according to our research, Klashi core YoY is not particularly well priced, not even 1h before the release.)
Table 1. Ten-day forecast scorecard on the identical four-model panel – pooled (MoM/YoY, core/headline).
Note: MAE and worst miss are in percentage points. Lower is better except for exact one-decimal accuracy. Best cells are highlighted.
Figure. Mean absolute error by CPI concept.
DIRECTION OF THE SURPRISE
The models correctly predict the direction of the surprise when a surprise occurs. We tested whether our models predict the direction of the surprise relative to Consensus and Kalshi, how reliably they do so, and under which conditions.
Large gaps between the models and markets forecasts are high-conviction directional signals. Conditional on a one-decimal Kalshi surprise, our bottom-up gets the direction right 71/98 times (72.4%) and our ML model 66/98 (67.3%). With a rounded ten-day gap of at least 0.2 percentage point, those rates rise to 12/12 (100%) and 17/19 (89%). The signal is visible at 30 days and remains present at 14 days and one hour before the CPI release.
Figure. Conditional directional hit rate
Note: The figure shows the direction win rate when the benchmark is surprised at one decimal, pooled across headline/core and MoM/YoY on the common ten-day panel. A directional win means the release lies on the model’s side of the benchmark; it does not require an exact forecast and does not by itself identify the best tradable strike.
Kalshi tends to converge toward the proprietary information set. From 30 to 14 days, when the gap between the model and Kalshi forecasts is at least five basis points, Kalshi moves closer to bottom-up in 57/64 cases (89.1%) and to ML in 58/76 (76.3%). The average distance closed is 9.7 and 6.5 basis points, respectively. This creates the possibility of a pre-release mark-to-market trade, not only a bet held to settlement.
Figure. Kalshi convergence.
Note: Panel A reports the frequency of a decline in absolute distance; Panel B reports the mean distance closed. The sample pools headline and core CPI, month on month and year on year, and requires an initial gap of at least 5 basis points.
TRADING THE MODELS ON KALSHI – BACKTEST AND INVESTMENT CONCLUSION
The reader can find all the details in the white paper. The strategy search covers a large set of no-reversal variants across model, CPI concept, horizon, strike, side and exit rule. Below we report the statistics of the trading strategies as well as the equity curve.
Table. Statistics of equal-weight strategy ensemble; $50,000 starting capital.
Figure. Capital evolution.
The conservative interpretation is that our models add information rather than merely replicate the market. Bottom-up is more accurate than the professional benchmark on the common sample and is materially more useful when the market is surprised. ML provides an independent signal, particularly for headline year-on-year inflation and for large model-market disagreements. The models are therefore most valuable not as isolated point estimates, but as a disciplined way to identify when the balance of risks embedded in market pricing is asymmetric. The evidence is consistent with a tradable edge: it appears well before the release, survives as the event approaches, and is present in a held-out trading sample after quoted prices and fees.
GASOLINE PRICES MODEL
We have developed a new model to forecast US retail gasoline prices. The model uses the national AAA price observed through each decision date together with the oil information then available. The model forecasts the calendar month average, the next Monday price, the next week price, and the end of the month price level. The production nowcast beats the Kalshi implied monthly average at seven of eight matched horizons.
For all details, please see the white paper.