AI Anomaly Detection for Finance Teams — deFacto Anomaly Detection

How dF Anomaly works in deFacto

A finance lead or FP&A admin sets the time range, the scenario to analyze, a confidence level, and optionally a set of exogenous business drivers, then triggers dF Anomaly from the planning surface. The tool analyzes historical data using the same Nixtla TimeGPT-2.1 engine that powers dF Foresight, then writes three companion scenarios into the model: a binary flag (1 = anomaly, 0 = normal), a lower confidence band, and an upper confidence band.

If the source scenario is Actual, the outputs are written as Actual_anomaly_99, Actual_anomaly_lo_99, and Actual_anomaly_hi_99 (assuming a 99% confidence interval). Finance reviews these alongside the original scenario in Excel or in the Power BI visual. For ongoing monitoring, dF Notify can subscribe to those flag tuples and send an email alert or surface the change in Power BI when a value trips the threshold. The whole chain — detection, write-back, and alert — runs on the connected deFacto model.

The same TimeGPT-2.1 engine as dF Foresight

dF Anomaly runs on the same Nixtla TimeGPT-2.1 model that powers dF Foresight. Azure deployments use TimeGPT-2 (Azure edition). For sovereign, private, or self-hosted environments, Google TimesFM 2.5 or Nixtla NeuralForecast 3.2 can be configured in its place. A consistent AI engine across forecasting and anomaly detection means your finance team works with predictable, auditable model behavior — not two separate black boxes.

Results write into scenarios finance already works in

The anomaly output writes only into scenarios that already exist in the deFacto model. If the companion scenarios are not set up, the tool warns and skips those outputs rather than creating ad hoc members. This is by design: anomaly results become part of the plan structure, reviewable and auditable the same way any other scenario is. Finance does not need a separate analytics tool or a new interface to inspect what the model found.

Exogenous drivers reduce false positives

dF Anomaly supports optional exogenous driver variables — for example, unit price or unit cost pulled from the fact table. Including business drivers means the detection accounts for known changes in the operating environment rather than flagging planned events as anomalies. A revenue movement driven by a scheduled price change looks different from an unexplained volume loss, and the model can distinguish between them when the relevant drivers are in scope.

Works best with adequate time series history

dF Anomaly performs best when the underlying data has sufficient history. Monthly series typically benefit from three or more years of actuals before the model produces reliable confidence bands. Short series return an insufficient-samples signal rather than a false positive. This is a quality gate, not a limitation: anomaly detection against thin history produces noise, not insight. The deFacto team can advise on data minimums for your specific scenario cadence and frequency during a demo or scoping engagement.

Recognized as a platform AI capability in the industry survey

The industry survey of planning platforms, July 2025, identifies anomaly detection as one of four AI capabilities in the deFacto platform profile — alongside intelligent forecasting, proactive notifications, and user support. These are discrete tools running on a connected model.

Where dF Anomaly fits in the deFacto AI model

dF Anomaly is one of four named AI offerings on the deFacto platform. dF Foresight handles time-series forecasting with scenario write-back. dF Anomaly detects drift. dF Notify alerts subscribed users when flagged tuples change. And conversational agents answer questions from the connected model and the product knowledge base.

All four run on the same connected deFacto model. That means anomaly flags, forecasts, and alerts all reference the same data your finance team plans against. For deployments that require a locally hosted model, the sovereign AI path supports Google TimesFM 2.5 and Nixtla NeuralForecast 3.2 without a cloud dependency.

See also: AI planning software hub · Compare forecasting platforms · Compare financial analysis platforms · Compare to Planful · Compare to Pigment