Databricks turns driver analysis and anomaly detection into SQL function calls

When an analyst calls ai_forecast, the series goes to a transformer-based time series foundation model, pretrained on a large and diverse collection of series ^[5]. Databricks says the model already recognizes common patterns such as seasonality and trends, so it needs no tuning ^[5]. Databricks maintains the model centrally ^[7].

The post's demo projects monthly NYC taxi ride volumes through the end of 2027 ^[12]. That series has a seasonal shape, which is the case Databricks describes. A series pushed around by a price change or a campaign is a different case. For that, ai_forecast has built-in holiday support and accepts external covariates, such as pricing changes or marketing spend ^[6]. For the demo to transfer to your workload, your series would need to look like the ones the model was pretrained on, or your covariates would need to explain what does not.

ai_detect_anomalies takes prepared input. In the example, taxi trips are aggregated into a daily revenue time series and passed to the function, which fits a model automatically and flags unusual days ^[9]. The aggregation, and its grain, stay with the analyst.

ai_top_drivers, one of the four new functions ^[14], is the cleanest design in the set. Its example compares January and February 2016 taxi fare revenue across time-of-day and weekday/weekend segments and ranks each segment by absolute and relative revenue impact ^[11]. Reporting both measures separates a large segment that moved a little from a small one that moved a lot ^[11].

Databricks says the upgraded ai_forecast is more accurate ^[3]. The post, in the text we reviewed, includes no figure and no baseline for that claim.

We think these functions fit ad-hoc questions well. Databricks says the queue for a data science team means "many questions never even get asked" ^[13]. A forecast that sets a budget is a different job, and it should be backtested against your own history before anyone relies on it.