mostlyright 0.14.0

Settlement truth layer for prediction markets. Raw weather observations, climate records, and forecasts with point-in-time query semantics.

MostlyRight

Settlement truth layer for prediction markets. Raw weather observations, climate records, and forecasts with point-in-time query semantics — for quants training Kalshi NHIGH/NLOW temperature models.

Install

pip install mostlyright             # SDK only (~2 deps: httpx, jsonschema)
pip install mostlyright[parquet]    # adds pandas + pyarrow for DataFrames

Quickstart

from mostlyright import MostlyRightClient

with MostlyRightClient() as client:
    obs = client.observations("NYC", from_date="2024-07-04", to_date="2024-07-04")
    # → list[Observation] with 30 raw METAR fields (temp_f, wind_speed_kt, ...)

Training workflow

Paired observations + NWS CLI settlement + optional forecasts, one row per settlement date. The primary training surface.

from mostlyright import MostlyRightClient

with MostlyRightClient() as client:
    # 5 years, joined against IEM MOS forecasts + Open-Meteo fallback.
    train = client.pairs(
        "NYC",
        from_date="2020-01-01",
        to_date="2024-12-31",
        include_forecast=True,
        as_dataframe=True,      # requires mostlyright[parquet]
    )
    # Columns: date, cli_high_f, cli_low_f (settlement truth),
    #          obs_high_f, obs_low_f (observed),
    #          fcst_high_f, fcst_low_f, fcst_model, fcst_issued_at, ...

Inference workflow

Live forecast signal, fetched directly from the upstream source — no MostlyRight infrastructure between the quant and the weather model. Column parity with the historical client.forecasts() output is the contract.

import mostlyright

live = mostlyright.live.get_forecast("NYC", model="NBS", source="iem")
# → pandas.DataFrame with the same columns as client.forecasts(...)
prediction = model.predict(live)

# Or Open-Meteo (seamless hourly):
live_om = mostlyright.live.get_forecast(
    "NYC", model="gfs_seamless", source="open_meteo"
)

Authentication

Dev mode requires no key. Production quants should set:

export MOSTLYRIGHT_API_KEY=<your-key>

Or pass explicitly:

import os
from mostlyright import MostlyRightClient

client = MostlyRightClient()  # reads MOSTLYRIGHT_API_KEY from env
# Legacy THERMINAL_API_KEY is also honoured for drop-in migration.

Point-in-time queries

Every method accepts an as_of watermark. Results are filtered so an inference loop on historical data sees only what was knowable at that UTC moment — climate records are withheld until the NWS CLI publication threshold has passed.

snap = client.snapshot("NYC", as_of="2024-07-04T18:00:00Z")
snap.observations             # filtered to [window_start_utc, as_of]
snap.climate                  # None until CLI publishes (~10h after midnight LST)
snap.version                  # DataVersion — reproducibility token

Data methods

Method	Returns	Purpose
client.observations(station, from_date, to_date)	METAR/SPECI observations	30 raw fields, up to 24/day
client.climate(station, from_date, to_date)	NWS CLI daily records	Kalshi settlement source
client.forecasts(station, ...)	IEM MOS runs	Discrete issued_at runs, per-hour rows
client.forecast_series(station, ...)	Open-Meteo hourly	Seamless hourly, metric/SI
client.pairs(station, from_date, to_date)	Settlement-date rows	Observations + climate + optional forecast, joined
client.snapshot(station, as_of)	DataSnapshot	All data available at a UTC moment
client.feature_catalog()	Feature metadata	Discovery surface for AI agents
mostlyright.live.get_forecast(station, ...)	pandas.DataFrame	Live upstream forecast

See docs/QUICKSTART.md for the longer walkthrough of each method, format options (toon, parquet, csv), and settlement window conventions.

Available stations

20 airports with active Kalshi NHIGH/NLOW markets. Codes accept 3-letter NWS form (NYC) or 4-letter ICAO (KNYC):

ATL  AUS  BOS  DCA  DEN  DFW  HOU  LAS  LAX  MDW
MIA  MSP  MSY  NYC  OKC  PHL  PHX  SAT  SEA  SFO

Station metadata (lat/lon, timezone, ICAO) is available at runtime:

stations = client.stations()             # list[StationInfo]
info = client.station("NYC")             # one station

Error handling

The SDK raises typed exceptions mapped from HTTP status codes:

from mostlyright.exceptions import (
    NotFoundError, ValidationError, RateLimitError,
    AuthenticationError, ForbiddenError, ServerError,
)

try:
    obs = client.observations("XYZ")  # unknown station
except NotFoundError as exc:
    ...
except RateLimitError as exc:
    wait_s = exc.retry_after

All exception classes subclass mostlyright.exceptions.TherminalError.

Schemas

Every data entity has a JSON Schema under specs/ on GitHub (observation.json, climate.json, forecast.json, forecast_series.json, snapshot.json, market_unified.json, ...). The SDK validates every response against these schemas before returning.

Changelog

Release notes live on GitHub Releases.

License

MIT — see LICENSE.