Changelog

v1.4.2 — 2026-03-31

Fix: parquet nullable columns: Observation parquet files now use Optional schema for nullable fields. Previously, the Go API inferred the parquet schema from the first row only and created non-nullable columns — null values were silently written as zero (0 for int, 0.0 for float), skewing downstream analysis. JSON was unaffected.
API: writeParquetTyped now scans all rows for the full column union and wraps nullable columns with parquet.Optional()
Python SDK: 7 new regression tests verifying nullable parquet columns produce NaN in DataFrames, not 0. No SDK code changes needed — PyArrow handles optional columns correctly.
213 unit tests + 45 integration tests

Integration test suite: 45 parametric E2E tests against the live API + AWC, covering every public function x every output mode (list, DataFrame, CSV, parquet)
CI publish gate: integration tests must pass before PyPI publish on tag push
Fix: parquet bytes→str: DataFrame columns from parquet now return str, not bytes (PyArrow types_mapper)
Fix: climate dates tz-naive: climate observation_date index stays tz-naive (calendar dates), not UTC
Fix: LOB numeric index: LOB ts column correctly detected as numeric epoch, not parsed as date string
Fix: empty DataFrame NaN edge case: empty result sets no longer misclassify index timezone
Fix: DataFrame immutability: _apply_datetime_index no longer mutates input DataFrame
CI security: pypa/gh-action-pypi-publish pinned to commit SHA, publish step gated on tag ref
206 unit tests + 45 integration tests

Domain-split modules: therminal.weather (WeatherHistory, WeatherLive) and therminal.markets (MarketsClient)
Canonical names: WeatherClient renamed to WeatherHistory, LiveClient renamed to WeatherLive
WeatherLive: Real-time METAR from AWC (aviationweather.gov), direct fetch with identical Observation schema to historical data
- Batch support: live.current(["NYC", "ATL", "MDW"]) makes one AWC request
- as_dataframe=True on current() and latest()
- Relative humidity computed via Magnus formula (native Celsius, cross-checked against IEM for 18 stations)
- Feels-like temperature via full NWS algorithm (wind chill + Steadman + Rothfusz + adjustments)
- Precipitation mapped from AWC precip field
- Peak wind gust/direction/time parsed from raw METAR PK WND remarks
- Go-matched rounding ensures training/inference parity
Auto-pagination: JSON responses transparently paginated, as_dataframe=True uses parquet for full datasets (no 50K JSON cap)
TherminalConfig: ~/.therminal.toml config file with 4-layer resolution (file < env vars < kwargs)
Typed models: Frozen dataclasses with dict-style access (obs["temp_f"], obs.get(), obs.temp_f)
Security hardening: guarded float() on all AWC numeric fields, direction validation, truncated unbounded strings
Python 3.11+ required (dropped 3.10)
TherminalClient facade preserved for backward compatibility
206 tests, 80% coverage

GET /api/v1/lob — Reconstructed order book snapshots from Kalshi and Polymarket delta data
Parameters: market, series, source (kalshi/polymarket), date/from/to, interval (1–3600s, default 60), levels (1–50, default 10)
Supports JSON, CSV, Parquet, and NDJSON output formats
Singleflight deduplication for identical concurrent requests
4 concurrent LOB requests max (semaphore-bounded)
Result caps: 50K JSON, 500K CSV/Parquet/NDJSON

client.lob() — LOB snapshots with full format support, DataFrame, save_path
LOBFeatures sklearn transformer — spread, mid, imbalance, depth, weighted_mid, pressure features with lookback aggregation
CLI: therminal lob MARKET --date 2026-03-01 --interval 60

?resolution=1min on /api/v1/observations returns 1-minute ASOS data from NCEI (DSI-6406)
~166M records, 2000–2026, 19 of 20 stations (NYC excluded — Central Park is not an airport ASOS)
Response fields: station_code, observed_at, temp_c, dewpoint_c, pressure1_inhg, pressure2_inhg, pressure3_inhg, precip_type, precip_in
?units= rejected with 400 for resolution=1min (temperatures always integer °C)
?type= silently ignored for resolution=1min
?format=csv|parquet and ?columns= work normally

New report type ncei_final (priority 2.5) on /api/v1/climate — GHCN-Daily CF6 first-published values from NCEI superghcnd diffs
Source tagged as ncei

Added resolution parameter to observations() — resolution="1min" for 1-minute ASOS data
CLI: therminal observations STATION --resolution 1min

?format=json|csv|parquet on all data endpoints (candles, observations, climate)
?columns=col1,col2 for column selection in CSV/Parquet output
Removed old /api/v1/download/* presigned URL endpoints
CSV streams with Content-Disposition header for browser download
Parquet uses Zstd compression

Added format and columns params to candles(), observations(), climate()
Removed download_parquet() (use format="parquet" instead)

Initial public release.

GET /health — API health check
GET /api/v1/series — List and retrieve event series
GET /api/v1/markets — List and retrieve prediction markets
GET /api/v1/candles — OHLCV candlestick data with forward-fill and aggregation
GET /api/v1/observations — Weather observations (METAR/SPECI)
GET /api/v1/climate — Daily climate reports
GET /api/v1/analysis/climate-gaps — Climate report gap analysis
GET /api/v1/download/observations — Bulk observation parquet download
GET /api/v1/download/candles — Bulk candle parquet download
GET /api/v1/download/climate — Bulk climate parquet download