AfyaSolar · Chronos-48M Climate — East Africa (v1)
A Chronos-Bolt (48M-parameter) time-series forecaster for daily and monthly climate variables across East Africa, packaged as two AutoGluon-TimeSeries predictors. It powers the AfyaSolar platform, which supports solar-powered health facilities; the app derives climate-hazard indices and solar-yield estimates from these forecasts.
| Base model | amazon/chronos-bolt-small (~48M params, Apache-2.0) |
| Variables (7) | ALLSKY_SFC_SW_DWN, T2M, T2M_MAX, T2M_MIN, PRECTOTCORR, WS10M, RH2M |
| Region | East Africa — 275-point 1° land grid |
| Training data | NASA POWER daily, 2000→present — dataset repo |
| Horizons | daily (H=30, freq D), monthly (H=12, freq MS) |
| Served model | ChronosZeroShot — fine-tuning gave no gain over zero-shot (see */leaderboard.csv) |
Repository layout
monthly/ AutoGluon TimeSeriesPredictor (H=12)
daily/ AutoGluon TimeSeriesPredictor (H=30)
model_card.json machine-readable metadata + validation scores
Usage
from huggingface_hub import snapshot_download
from autogluon.timeseries import TimeSeriesPredictor
path = snapshot_download("afyalink/afyasolar-chronos-48m-climate-ea-v1")
predictor = TimeSeriesPredictor.load(f"{path}/monthly")
# predictor.predict(your_TimeSeriesDataFrame)
Compatibility: trained with AutoGluon 1.6.1 on Linux. Load with a matching
version. On Windows or an older AutoGluon, load with require_version_match=False
and map pathlib.PosixPath = pathlib.WindowsPath before loading (the pickles
contain Linux paths). Zero-shot Chronos is stable across these minor versions.
Evaluation
AutoGluon 3-window rolling backtest, metric WQL (lower is better):
| horizon | SeasonalNaive | ZeroShot | FineTuned |
|---|---|---|---|
| monthly | 0.253 | 0.162 | 0.163 |
| daily | 0.110 | 0.053 | 0.053 |
Chronos beats the seasonal baseline by ~36% (monthly) / ~52% (daily); fine-tuning did not improve on zero-shot, so zero-shot is served.
License
Apache-2.0 (inherits Chronos-Bolt). NASA POWER source data is freely available.