TSFM.ai developer documentation.
Multiple pages, one contract. API, MCP, and CLI are aligned on the same schema so teams can move from manual calls to production automation with zero drift.
Learn
Use cases for time series foundation models
TSFMs are general-purpose models that can be applied to forecasting, anomaly detection, classification, and imputation across virtually any industry that works with temporal data.
Supported tasks
Forecasting
/v1/forecastThe primary use case. Given historical observations, predict future values with uncertainty estimates. TSFMs produce point forecasts, quantile intervals, and full predictive distributions depending on the model.
Examples
- Demand planning — forecast next-week unit sales per SKU to optimize inventory allocation
- Energy load forecasting — predict hourly electricity consumption for grid balancing
- Financial projections — forecast revenue run rate with confidence intervals for board reporting
- Capacity planning — predict server request volumes to right-size infrastructure
Recommended models: All models on TSFM.ai support forecasting. Chronos-2 and TimesFM are popular starting points.
Anomaly detection
/v1/detect-anomaliesIdentify unusual observations that deviate significantly from expected patterns. The model learns what 'normal' looks like from the series history and flags data points that fall outside the expected distribution.
Examples
- Infrastructure monitoring — detect CPU spikes, memory leaks, and network anomalies in real-time
- Fraud detection — flag unusual transaction patterns in payment systems
- Quality control — identify sensor readings that indicate manufacturing defects
- Security — detect unusual access patterns in application logs
Recommended models: MOMENT is purpose-built for multi-task including anomaly detection. Toto excels at observability signals.
Classification
/v1/classifyCategorize time series into predefined classes based on their temporal characteristics. Useful for pattern recognition, regime detection, and automated labeling of signal types.
Examples
- Regime detection — classify market conditions as trending, mean-reverting, or volatile
- Sensor type identification — automatically categorize IoT sensor streams by signal type
- Activity recognition — classify user behavior patterns from interaction timelines
- Failure mode classification — identify which type of equipment failure a vibration signal indicates
Recommended models: MOMENT supports classification from a single checkpoint. Other models can be fine-tuned for classification tasks.
Imputation
/v1/imputeFill in missing values in time series data while preserving statistical properties. TSFMs understand the expected patterns and can reconstruct gaps that are consistent with the surrounding context.
Examples
- Sensor data gaps — fill in missing readings from temporarily offline sensors
- Survey data — interpolate missing responses while maintaining distributional properties
- Historical data cleaning — reconstruct corrupted or incomplete historical records
- Multi-source alignment — fill gaps when merging data from sources with different reporting schedules
Recommended models: MOMENT handles imputation natively. For other models, a forecast-based approach can approximate imputation.
Industry applications
TSFMs are being adopted across industries. Here is how organizations are applying them today.
Retail and e-commerce
Demand forecasting
Predict unit sales per SKU at store level for replenishment and markdown optimization
Promotion impact
Model the lift from planned promotions using future covariates
New product launch
Zero-shot forecasting for items with no sales history using transfer from similar categories
Energy and utilities
Load forecasting
Predict hourly and daily electricity demand for grid operations and trading
Renewable generation
Forecast solar and wind output to balance supply and demand
Anomaly detection
Identify unusual consumption patterns indicating equipment issues or theft
Finance and insurance
Revenue forecasting
Project financial metrics with prediction intervals for scenario planning
Risk modeling
Quantify tail risk using probabilistic forecast distributions
Claims volume
Predict insurance claim volumes by category for reserve allocation
Infrastructure and DevOps
Capacity planning
Forecast request volumes and resource utilization for auto-scaling
Incident detection
Detect anomalous latency, error rate, and throughput patterns
Cost forecasting
Predict cloud spend trends for budget planning and optimization
Healthcare and life sciences
Patient volume
Forecast emergency department and clinic visit volumes for staffing
Clinical monitoring
Detect anomalous vital sign patterns in continuous patient monitoring
Drug demand
Predict pharmaceutical demand by region to prevent stockouts
Manufacturing and supply chain
Production planning
Forecast demand upstream to optimize production schedules
Predictive maintenance
Detect degradation patterns in equipment sensor data before failures
Logistics optimization
Predict shipping volumes and transit times for route planning
When TSFMs may not be the best fit
TSFMs are powerful but not universally optimal. Consider alternatives when:
- You have a single, well-understood series with decades of clean history — a tuned ARIMA or ETS may suffice and be cheaper to run.
- Your data has strong domain-specific structure that requires custom feature engineering — tree-based models (LightGBM, XGBoost) may capture engineered features better.
- You need sub-millisecond latency — even fast TSFMs add 50-100ms of GPU inference time. Simple statistical methods can run in microseconds.
- Your series is fundamentally unpredictable (pure random walk) — no model, classical or foundation, will produce meaningful forecasts.
Choosing a model
Use a decision framework to select the right TSFM based on your latency, cost, and accuracy requirements.
Selection guide
Quickstart
Ready to try it? Make your first forecast API call in under 5 minutes.
Get started