General Forecast

Predictors on tabular and time series foundation models. One manifest, gf.yaml, runs the same predictor on your machine and hosted.

• Website: https://generalforecast.com
• Docs: https://docs.generalforecast.com/

pip install generalforecast

A predictor from a CSV

In a directory with a CSV, gf init asks for the target, the model, the id column, the inputs and which rows are queries, then writes gf.yaml.

$ cd churn && gf init
Target       churned
Model        general-forecast/tabicl-v2-classifier
Id           customer_id
Inputs       spend, plan, signups
Queries      212 rows with no target

 Initialized predictor churn in ./gf.yaml

gf check resolves the schema against the context and pins the model in gf.lock.

$ gf check
    Resolved churn  classification, 3 inputs, 1,840 context rows, 212 query rows
      Locked gf.lock in 0.21s

gf eval scores the held-out rows.

$ gf eval
    Resolved churn  classification, 1,840 context rows, 212 query rows
   Predicted 212 rows in 0.19s

             accuracy   0.934
             log loss   0.187
             auroc      0.961

gf serve runs the predictor on this machine. The first run pulls the model's weights into ~/.cache/huggingface.

$ gf serve
     Pulling general-forecast/tabicl-v2-classifier  (1.2 GB) to ~/.cache/huggingface
     Loading general-forecast/tabicl-v2-classifier (cached)
    Compiled 12 programs in 6.8s
     Serving churn at http://127.0.0.1:8000

curl -s http://127.0.0.1:8000/v1/projects/local/predictors/churn/predict \
  -H 'content-type: application/json' \
  -d '{
    "api_version": "gf/v1",
    "project": "local",
    "request_id": "r-1",
    "resource": {"kind": "predictor", "name": "churn"},
    "payload": {"queries": [{"customer_id": 9001, "spend": 42.0, "plan": "pro", "signups": 3}]}
  }'

The same predictor from Python:

import generalforecast as gf
import pandas as pd

df = pd.read_csv("new_customers.csv")
scored = gf.Predictor("./churn").predict(df)  # columns: customer_id, label, p_0, p_1

gf push deploys it. gf predict and gf.Predictor("churn") then score rows against the hosted revision.

$ gf push
   Uploading churn revision 7  snapshot, 1,840 rows, 3.1 kB
    Deployed churn revision 7 to production in 41s

Estimators

Classifier, Regressor and Forecaster have the sklearn shape and take a catalog model. Weights pull on first use; download=False or GF_OFFLINE=1 refuses the download and names the gf model pull command instead.

import generalforecast as gf

clf = gf.Classifier(model="general-forecast/tabicl-v2-classifier")
clf.fit(X_context, y_context)
clf.predict_proba(X_queries)

fc = gf.Forecaster(model="general-forecast/chronos-2", horizon=7)
fc.fit(history)  # columns: date, sku, units
fc.forecast()  # quantiles 0.1, 0.5, 0.9 per sku per step

gf model list prints the catalog with each model's limits on context rows, queries and inputs.

Contributing

We welcome and value any contributions and collaborations. For more specific questions, please contact hello@generalforecast.com.

License

Licensed under the Apache License 2.0.

If you use General Forecast, please cite our work:

@misc{generalforecast2026,
    title={General Forecast: A unified high-throughput inference engine for structured data foundation models},
    author={Tejas Prabhune and Ayushi Batwara},
    year={2026},
    publisher={GitHub},
    howpublished={\url{https://github.com/generalforecast/gf}}
}