v0.9.0 #903

senbong (gee-senbong) · 2026-09-21T05:33:00Z

senbong (gee-senbong)
Sep 21, 2026
Maintainer

Highlights

Asynchronous jobs (client.jobs) — long-running work can now run server-side without blocking. jobs.forecast(), jobs.cross_validation(), jobs.finetune(), jobs.detect_anomalies(), jobs.simulate(), jobs.explain() and jobs.execute_step() each return a Job handle with .wait(), .status and .cancel(), mirroring the blocking method of the same name — note that jobs.detect_anomalies() mirrors detect_anomalies_online(), i.e. it runs online detection. Job, JobStatus, JobError, JobTimeoutError, JobCancelledError and ApiError are exported from the package root (feat: add async forecast, cross validation & finetune method #847, [FEAT] Add submit_anomaly_detection_job for asynchronous anomaly detection #879, [FEAT] Unify async job handling; add submit_simulate_job and submit_explain_job #883, [REFACTOR] Move async jobs into a client.jobs namespace and a nixtla/jobs package #887).
simulate() — draws temporally correlated sample paths per series, in long format with a sample_id column. multivariate=True keeps paths coherent across series ([FEAT] Simulate and explain #852).
explain() — model-independent feature importance weights, via "granger" or "transfer_entropy". These describe predictive relationships, not causal ones ([FEAT] Simulate and explain #852).
list_models() — returns the model names your API key may use. Needs the GET /v2/models route, which is not deployed yet; until it ships the call raises ApiError with status 404 ([FEAT] Add list_models to list the models available to an API key #881).
nixtla.snowflake helpers — a package for driving the Snowflake TimeGPT integration from Python, exposing detect_anomalies, load_actuals and to_anomalies_df (feat: snowflake demo helpers and collateral + unpin claude #870).
Start datetime information — requests now carry start-datetime metadata for each series ([FEAT] Start datetime information #859).
Breaking: detect_anomalies_online() gained finetuned_model_id and model_parameters mid-signature, to match the parameter order of forecast() and cross_validation(). Callers passing arguments positionally past finetune_loss must switch to keywords ([FEAT] Unify async job handling; add submit_simulate_job and submit_explain_job #883).
Deprecation ahead of 1.0: the anomaly-detection methods are being renamed and detect_anomalies() changes meaning. Batch detect_anomalies() is removed in 1.0 with no drop-in replacement; detect_anomalies_online() takes over the name, unchanged in behavior. The Snowflake anomaly path (nixtla.snowflake.detect_anomalies(), NIXTLA_DETECT_ANOMALIES and its UDTF) makes the same move, so h and detection_size become required PARAMS entries. Every affected entry point now emits a FutureWarning ([CHORE] Warn about the 1.0 anomaly-detection rename #896).
Cross-validation correctness fix — actual values returned in the cross-validation output were incorrect and are now right; re-verify any evaluation metrics computed with v0.8.0 ([FIX] Incorrect actual values in cross-validation output #866).