Kalshi CPI Expectations

An event-study pipeline that recovers a minute-level distribution of expected US CPI inflation from Kalshi bracket markets and estimates its response to CPI announcement surprises.

The repository separates two claims that are easy to conflate:

• The software is publicly reproducible. make demo builds synthetic inputs, runs the same panel, distribution, regression, diagnostic, and figure code, and requires no API key or proprietary file.
• The empirical estimates require licensed Bloomberg inputs. Bloomberg's release calendar and consensus data are not distributed. A user with an authorized export can reproduce the research results locally.

Research question

For each CPI release, Kalshi's threshold contracts imply exceedance probabilities, such as (P(CPI > 0.3)). Adjacent thresholds recover probability mass over outcome bins. The pipeline uses that distribution to calculate (E[CPI]) and (Var(CPI)), then estimates

[ \Delta E_i^{(h)} = \alpha_h + \beta_h Surprise_i + \varepsilon_i, ]

where the baseline is exactly five minutes before publication and (h \in {5,10,15,30,60,90,120,240}) minutes.

What the empirical results show

The checked-in aggregate results cover 40 CPI releases in the final consistent sample. The estimated response of expected CPI is positive and becomes clearer at longer horizons. At 120 minutes, (\hat\beta=0.0212) with (p=0.0013). These estimates should be read with the data-quality limitations below: about 27% of events have essentially no measured movement at five minutes, which may represent either genuine non-reaction or stale quotes.

The repository does not claim that the existing random-day placebo validates identification. The stored panel is built around release windows, so it does not contain suitable non-announcement days. The supported public run reports the pre-trend diagnostic and leaves the random-day placebo disabled.

Selected portfolio figures are checked in under reports/figures/; the full event-level figure set is generated locally.

The reviewed term paper is available at paper/inflation-expectations-cpi.pdf. Its title page omits course administration and student identifiers.

Quick start: fully public demo

Requirements: Python 3.11 or 3.12, make, and roughly 1 GB of free space.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements-lock.txt
make test
make demo

The demo writes only to build/demo/. Its events, surprises, quotes, and results are deterministic synthetic data and have no empirical meaning. The complete run produces a compact panel, event-level regression data, diagnostics, tables, and figures.

Reproduce the licensed empirical analysis

1. Obtain an authorized Bloomberg Excel export with columns Event, Date Time, Actual, Surprise, Surv(M), and Prior.
2. Save it locally as data/private/bloomberg_events_data.xlsx. That directory is ignored by Git and must never be committed. A different location can be selected with BLOOMBERG_EVENTS_FILE=/absolute/path/to/export.xlsx.
3. Install dependencies as above. requirements-lock.txt records the verified environment; requirements.txt allows compatible newer versions for development.
4. Run:

python scripts/01_parse_bloomberg.py
python scripts/02_discover_markets.py
python scripts/03_download_candles.py
python scripts/04_process_panel.py
python scripts/05_prepare_regression.py
python scripts/06_robustness_checks.py --pretrend-only
python scripts/07_make_plots.py

Kalshi's historical public endpoints do not require a key in this implementation. API availability and historical coverage can change, so exact raw-data retrieval is not guaranteed indefinitely. Existing downloads are incremental.

Pipeline

Step	Script	Purpose
00	00_generate_demo_data.py	Create public synthetic inputs
01	01_parse_bloomberg.py	Parse the private licensed event export
02	02_discover_markets.py	Discover CPI markets and brackets
03	03_download_candles.py	Download public one-minute quote candles
04	04_process_panel.py	Build compact ([-240,+240])-minute panels
05	05_prepare_regression.py	Infer distributions and estimate OLS models
06	06_robustness_checks.py	Run pre-trend diagnostics; experimental placebo code is disabled by default
07	07_make_plots.py	Generate figures and liquidity diagnostics

Step 04 has an explicit --full-history option. It can produce a panel hundreds of megabytes in size and is not necessary for the main event study.

Sample construction

An event enters every reported horizon only when it has:

1. an exact observation at (t=-5);
2. at least three non-placeholder brackets at baseline;
3. finite inferred expectation and variance; and
4. valid estimates at all eight horizons.

A quote with bid 0 and ask 100 is treated as an empty-order-book placeholder. Quotes are forward-filled between observed updates. Consequently, a flat path can mean either stable beliefs or no new quote activity; the data cannot identify which explanation applies.

Repository map

config/       Series configuration
docs/         Data, methodology, and API notes
scripts/      Numbered pipeline entry points
src/          Reusable parsing, panel, and distribution functions
tests/        Unit tests used by GitHub Actions
reports/      Reviewed aggregate empirical tables
build/        Ignored synthetic demo output
data/         Ignored licensed, raw, and intermediate local data

See METHODOLOGY.md for the full method and docs/DATA_AND_REPRODUCIBILITY.md for the publication boundary.

Limitations

• Bloomberg consensus data cannot be redistributed with this repository.
• Sparse trading and forward-filled quotes can attenuate response estimates.
• The strict baseline selects events with usable pre-release liquidity.
• The current data do not support a credible random non-event-day placebo.
• Bin endpoints require tail assumptions; results are market-implied measures, not direct survey forecasts.

License and citation

Code is released under the MIT License. That license does not apply to Bloomberg data or third-party market data. If you use the project, cite the repository URL and the version or commit hash used.