beita6969 / scienceclaw-benchmark-for41

Probabilistic aquatic forecasts (oxygen and temperature, CRPS)

Task

• Input: history (n, L, 3) daily means (channel 0 oxygen mg/L, 1 temperature degC, 2 chlorophyll-a ug/L, not forecast), the last column being each item's reference_date; NaN = not observed or withheld; site id and type may be given.
• Deliverable: a dict of four float arrays (n, 30): oxygen_mu, oxygen_sigma, temperature_mu, temperature_sigma for the days reference_date + 1 .. + 30. All finite, sigma > 0, oxygen mu in [0, 30] mg/L, temperature mu in [-5, 45] degC (this catches kelvin), sigma at most 50.
• Quality: closed-form CRPS of N(mu, sigma^2) at the observed (item, day) pairs (unobserved days are not scored), averaged per variable; primary = equal-weight mean of the oxygen and temperature CRPS, lower is better. The two have different units, so report both (scilib.aquatics.score(pred, obs), crps_normal). Reference: day-of-year climatology (mean and sd within +-7 days of each target day of year, aquatics.doy_window_forecast).

Routes

• scilib.aquatics.fit_predict(history, reference_date, lam=0.75, smult=1.0) (operator aquatic_oxygen_temperature_forecast_30d): smooth day-of-year profile, then per lead day a ridge regression of the anomaly on the recent anomaly (mean of the last 3 and last 30 days), the site fit blended with a pooled fit by lam, sigma from the residual sd times smult. lam and smult moved the mean CRPS by only a few percent in backtests; do not over-tune them. Items with fewer than 60 observed days of a variable fall back to the climatology.
• Checks: aquatics.backtest(history, reference_date, n_origins=3, offset=30) is a rolling-origin test that needs only history; data_report shows coverage and age of the last observation; seasonal_profile, history_dates are building blocks. Operators aquatic_climatology_forecast_30d, normal_forecast_crps_score.
• Pretrained blend: aquatics.fit_predict(..., pretrained=w) mixes every mu and sigma with Chronos-2 (median as mu, (q75 - q25)/1.349 as sigma; scilib.tsfm, asset chronos_2), w in [0, 1]; w = 0 makes no GPU call. Check tsfm.available() and scienceclaw_tools(operation=weights). Pretraining overlap with public time-series corpora cannot be ruled out; state it.

Pitfalls

• CRPS punishes both overconfident and needlessly wide sigma; check calibration in a backtest rather than only the mean error. The remaining error at long leads is mostly weather-driven temperature shocks.
• Respect temporal causality: forecast each item from history that ends at its own reference date. If one site appears with several reference dates, a later history can contain an earlier item's forecast window; never use it for the earlier date, and keep withheld (NaN) windows as NaN, not zero.
• A backtest window at the end of a history can be partly withheld, so its score is a noisy guide.
• Keep units: temperature in degC (not K), oxygen in mg/L; clip mu and sigma to the ranges above.