ChakraTS and ChakraTab are pretrained large foundation models that forecast any time series and predict any tabular data with one API call. No training or fine-tuning required.
Pretrained frontier models behind one API.
| Benchmark | 1st on fev-bench by win rate, the leading time series benchmark |
|---|---|
| vs classical | Better on 98 of 100 tasks than seasonal naive, ETS, ARIMA, LightGBM, CatBoost |
| Speed | Seconds per call, up to 256 series and 2,048 steps ahead |
| Takes | The history of anything measured over time |
| Returns | The next steps ahead, each with a range |
| Industries | EnergyRetailHealthcareFinanceClimate |
| Benchmark | Top 5 on TabArena, the leading benchmark for structured data |
|---|---|
| vs classical | Better than tuned gradient boosting on most tables, no feature engineering |
| Speed | Seconds to minutes per table, classification and regression |
| Takes | A table with the answers you have, plus the rows to answer |
| Returns | A prediction for every new row, with probabilities |
| Industries | HealthcareFinanceTechIndustrialsLife sciences |
Against the methods teams use today, on public datasets.
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Grid load, day ahead | ERCOT daily system load | ChakraTS | 24% lower error |
| Day-ahead power prices | German market, EPF | ChakraTS | 60% lower error |
| Wind farm output | 10-minute turbine power, KDD Cup 2022 | ChakraTS | 40% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Product demand | M5 daily unit sales by item | ChakraTS | 15% lower error |
| Grocery orders | Rohlik daily orders by warehouse | ChakraTS | 31% lower error |
| Campaign response | Marketing campaign, 2,240 customers | ChakraTab | 14% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Company bankruptcy | Polish companies, 5 years | ChakraTab | 87% lower error |
| Company bankruptcy | Taiwanese companies, 1999 to 2009 | ChakraTab | 20% lower error |
| Macro indicators | FRED-MD, 120 monthly US series | ChakraTS | 10% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Maternal health risk | Clinical readings, 1,014 patients | ChakraTab | 18% lower error |
| COVID cases | UK nations, daily new cases | ChakraTS | 20% lower error |
| Blood donation | Transfusion service centre, 748 donors | ChakraTab | 10% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Truck component failure | Scania APS system, 76,000 trucks | ChakraTab | 15% lower error |
| Transformer temperature | ETT hourly oil temperature | ChakraTS | 19% lower error |
| Contaminant detection | Hazelnut spread production line | ChakraTab | 79% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Customer churn | Telco subscribers, 5,000 accounts | ChakraTab | 15% lower error |
| Cloud query volume | Amazon Redshift, 5-minute | ChakraTS | 41% lower error |
| IT operations metrics | BizITObs, 5-minute | ChakraTS | 41% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Passenger satisfaction | Airline survey, 130,000 passengers | ChakraTab | 19% lower error |
| Highway traffic | Seattle loop detectors, 5-minute | ChakraTS | 30% lower error |
| Road speeds | Shenzhen taxi GPS, hourly | ChakraTS | 40% lower error |
| Use case | Data | Model | vs best classical method |
|---|---|---|---|
| Protein structure | Physicochemical properties, 45,000 decoys | ChakraTab | 16% lower error |
| Biodegradability | QSAR, 1,055 chemicals | ChakraTab | 14% lower error |
| Star and galaxy classification | Sloan Digital Sky Survey DR17 | ChakraTab | 16% lower error |
Error reduction against the best classical method on each dataset. Time series: seasonal naive, AutoETS, AutoARIMA, a statistical ensemble, LightGBM and CatBoost, fev-bench protocol. Tables: tuned and ensembled CatBoost, LightGBM and XGBoost, TabArena protocol.
Full results on GitHub coming soon.
Measured on industry-leading open benchmarks.
Set up time with our founder. We work through your use case with you, on your data, and build what your product needs.
