Skip to content

Configuration reference

Every key in config.toml, its type, default, validation rule, and effect.

Configuration is loaded by config.py into frozen dataclasses with explicit validation — an unknown value fails at load with a ConfigError naming the key, not at 3 a.m. inside a blender. Start from config.example.toml.

--config PATH selects the file; the default is config.toml in the working directory.

Only [station] and [forecasts] are mandatory

Every other section has working defaults. A first config is about twelve lines — see Getting started.

Environment variables

There is exactly one environment-variable mechanism in the entire codebase: a value of the form "$VARNAME" in [truth_qc] synoptic_token is read from the environment at use time.

[truth_qc]
synoptic_token = "$SYNOPTIC_TOKEN"

No other setting consults the environment. Everything else is in the file, which is what makes the config_fingerprint a complete description of how a run was configured.

[station] — where truth comes from

Key Type Default Meaning
db_path path required the ambientweather2sqlite SQLite file
timezone IANA name required defines the local day for daily products and DST handling
latitude float required decimal degrees, north positive
longitude float required decimal degrees, east positive
elevation_m float required metres; used for pressure reduction and neighbour lapse adjustment
immutable bool false open the DB with ?immutable=1

immutable must be false for a live database

immutable = true is only safe for a static snapshot file. A live WAL database opened immutable will return stale or inconsistent reads. The default is the safe one.

[station.columns] — DB column → canonical channel

Maps your station's column names onto the channels the pipeline understands. Defaults cover a standard AmbientWeather unit:

[station.columns]
outTemp = "temp"
outHumi = "humidity"
avgwind = "wind_speed"
gustspeed = "wind_gust"
eventrain = "rain_counter"
AbsPress = "pressure_station"

Canonical channels: temp, humidity, wind_speed, wind_gust, rain_counter, pressure_station.

[station.units] — channel → source unit

[station.units]
temp = "degF"
humidity = "pct"
wind_speed = "mph"
wind_gust = "mph"
rain_counter = "inch"
pressure_station = "inHg"

Accepted units: degF, degC, pct, mph, ms, inch, mm, inHg, hpa. Everything is normalized to metric on ingest.

rain_counter must be the event counter

The pipeline expects a monotone counter that resets between events, and computes reset-aware increments from it. A field that already reports hourly accumulation is a different quantity — see Methods: precipitation §2.

Run grounded-weather-forecast qc after editing either table. A wrong mapping shows up immediately as a channel that is 100% null.

[forecasts] — where the providers' opinions come from

Key Type Default Meaning
db_path path required the omni-weather-forecast-apis archive
immutable bool false as above — keep false while the collector writes
sources list of str [] allowlist of source slugs; empty means every source discovered
max_forecast_age_hours float > 0 12.0 staleness cap: a source older than this is unavailable at that snapshot
exclude list of "source:variable" [] null one variable of one source without excluding the source

max_forecast_age_hours materially determines which sources count as available for 6-hourly providers, and therefore what the blend is actually averaging. Raising it admits staler forecasts; lowering it thins coverage. Check its effect against the correlation report rather than guessing.

[forecasts.max_lead_hours]

Per-source lead cap in hours. Rows beyond the cap are dropped at matrix build, trimming a provider's degraded horizon tail.

[forecasts.max_lead_hours]
stormglass_sg = 120

[dataset]

Key Type Default Meaning
dir path data where parquet outputs go
min_hour_coverage fraction in \([0,1]\) 0.8 minute coverage required for an interval-mean hourly truth
min_day_coverage fraction in \([0,1]\) 0.8 minute coverage required for a daily truth, against the day's actual length
pop_threshold_mm float > 0 0.254 the "measurable precipitation" threshold defining PoP — 0.01 inch
precip_reset_fraction fraction 0.5 a counter drop below this fraction of the prior value is a genuine reset; a smaller dip is noise

pop_threshold_mm changes what every PoP number in the system asserts. It is a convention, not a constant of nature.

[qc] — station quality control

Metric units throughout. Defaults exist for every standard channel; override per channel.

[qc.bounds]
temp = [-40.0, 55.0]

[qc.max_step]
temp = 5.0

[qc.flatline_minutes]
temp = 180
Sub-table Type Meaning
bounds [low, high] outside → OUT_OF_BOUNDS
max_step float > 0 per-minute rate limit for the two-sided spike rule; scales with the actual sample gap
flatline_minutes int > 0 run of bit-identical values that counts as a stuck sensor

Flagged samples become NULL, never corrected. See Methods: truth and QC §1.

[provider_qc] — plausibility QC on forecast values

Applied before grounding, so a single bad provider value cannot poison the fit. Truth is never consulted here, so it introduces no leakage.

Key Type Default Meaning
enabled bool true master switch
mad_k float > 0 5.0 scaled-MAD multiple for the cross-source outlier test
min_sources int > 0 4 providers required before the cross-source test runs at all
bounds table of [low, high] per-variable defaults absolute physical bounds
cross_source_variables list of str temp/dew/humidity/pressure/precip which variables get the outlier test
min_deviation table of float per-variable defaults absolute deviation floor

A value is nulled only when it exceeds both mad_k scaled MADs and the min_deviation floor.

Never add a variable to cross_source_variables without a min_deviation

On a dry day the cross-source median and MAD are both zero, so mad_k × MAD is zero and the MAD test alone flags any lone nonzero value. Without a floor, a variable added here will null all light precipitation. Always set both together. See Methods: precipitation §4.

Gusts and PoP are deliberately excluded from the cross-source pass: a gust disagreeing with its peers is often correct, and that is the point of a gust field.

[backfill.open_meteo] / [backfill.dynamical]

Key Type Default Meaning
models list of str per provider models to fetch
start_date date — first date to backfill
publication_lag_hours float > 0 6.0 dynamical only — fetched_at = init + lag
max_lead_hours float > 0 48.0 dynamical only — cap on backfilled lead

publication_lag_hours is a leakage control, not a performance knob. A model cycle initialized at 00Z is not published until hours later; pretending it was visible at init would inflate short-lead skill with information nobody had.

[ensembles]

Key Type Default Meaning
models list of str — Open-Meteo ensemble models to poll; empty disables ingestion
variables list of str temp_c, dew_point_c, wind_speed_ms, wind_gust_ms, pressure_sea_hpa, precip_mm which variables to reduce

Open-Meteo keeps only the latest run's members, so run ingest-ensembles at least once per model cycle or the members are gone. build-dataset must run after each ingest before the ens__* features are visible to backtests and serving.

[backtest]

Key Type Default Meaning
initial_train_days int > 0 90 archive span required before the first fold origin
step_days int > 0 7 spacing between rolling origins
rolling_window_days int > 0 180 training window length when --window rolling

initial_train_days + step_days is the minimum archive age before backtest produces anything at all. With defaults that is 97 days — the "no rolling-origin folds" message means you are not there yet.

[predict]

Key Type Default Meaning
selection str skill_per_slice how methods are chosen per slice
history_path path data/predict_history.parquet where served forecasts are appended for self-verification
minutely_tau_hours float > 0 3.0 anchor decay for the minutely fallback path
quantile_recalibration str or per-product table none one of none, pit, cqr

history_path is what self-verification reads. Point it somewhere durable — losing it loses the only measurement of serving-path drift.

minutely_tau_hours applies only when no minutely evidence exists. Once a minutely leaderboard has promoted constructions, serving uses those per (variable, bucket) and this value is inert. See Methods: combination §8.

quantile_recalibration may be a single string or a per-product table:

[predict.quantile_recalibration]
hourly = "none"
daily = "cqr"
minutely = "none"

[predict.methods] — per-slice pins

[predict.methods]
"hourly.temp_c" = "gbm"

A pin overrides the leaderboard. Useful for reproducing an old serving decision or isolating a method in production; it bypasses the promotion gate, so it is a deliberate override of the system's own evidence rather than a tuning knob.

[promotion] — the statistical gate

Key Type Default Meaning
rule mcs | legacy | seq_mcs mcs which promotion gate to apply
alpha float > 0 0.1 level for the MCS, the e-process threshold \(\ln(1/\alpha)\), and FDR control
live_gap_factor float > 0 1.5 demote when realized MAE exceeds this multiple of backtest MAE
min_live_n int > 0 24 scored live cases required before demotion can fire
report_gap_threshold float > 0 0.15 relative gap above which a slice is listed under "Blocked promotions"
mcs_bootstrap int > 0 500 moving-block bootstrap replicates
mcs_block_length int ≥ 0 0 block length; 0 means \(\approx n^{1/3}\)

[promotion.references]

Per-variable overrides of the default reference set (best_provider, equal_weight, damped_grounded_equal_weight). The leaderboard computes columns for the union of defaults and overrides, so a pinned skill_vs_equal_weight column never silently changes meaning.

Full treatment: Methods: model selection.

[truth_qc] — neighbour cross-checks

Key Type Default Meaning
synoptic_token str — Synoptic Data token; "$VAR" reads the environment
radius_km float > 0 25.0 neighbour search radius
elevation_band_m float > 0 300.0 neighbour elevation band around the station
lapse_k_per_km float > 0 6.5 lapse rate used to adjust neighbour temperatures to station elevation
drift_statistic snht | pettitt snht change-point test
gate_fitting bool false null temperature labels under a latched drift verdict

NWS METAR neighbours need no token; Synoptic is the optional richer source.

lapse_k_per_km = 6.5 is the standard environmental lapse rate, not your mountainside's. A strong inversion reverses its sign — if your station sits in a cold-air pool, this default will systematically misjudge neighbour agreement.

gate_fitting is off by default, deliberately

Turning it on lets a drift verdict delete temperature truth. A wrong quarantine silently destroys the labels everything is scored against. Turn it on only after you have watched the verdict behave for a while.

[reports] and [artifacts]

Key Type Default
[reports] dir path reports
[artifacts] dir path artifacts

See Outputs for what lands in each.

Keeping config.toml out of git

config.toml holds your station's real coordinates and local paths. config.example.toml is the committed template; copy it, edit the copy, and keep the copy untracked.