Models

Every forecasting model, in one place.

Use pretrained foundation models for instant forecasts, fine-tune them on your data, build custom architectures, or bring your own — and compare them all on the same benchmark.

Model library listing custom XGBoost, LSTM, and Random Forest models alongside discovered tournament winners — The model library — custom builds, tournament winners, and uploads, side by side.

Foundation models

A pretrained forecaster, ready in one call.

Deploy a foundation model zero-shot for an immediate forecast without a training run. When you need more specialization, fine-tune a lightweight adapter on your own segment while the foundation weights stay frozen.

Zero-shot deployment with quantile forecasts
Fine-tune an adapter: cheap to train, cheap to store, easy to swap
Platform and supported community backbones in one catalog

Foundation model catalog with pretrained forecasters by task, parameter count, and license — Browse foundation models by task, parameter count, and framework.

Custom & BYOM

Build your own. Or bring one you already trained.

Engineer a custom architecture on-platform, or upload a model trained elsewhere and register it in the catalog after validation. Custom and BYOM models train and race exactly like everything else.

Visual model builder — LSTM, GRU, Transformer, and classic learners
BYOM upload with automatic validation before it's ever trusted
Versioned trainings and artifacts — nothing lives only on a laptop

Model studio wizard for composing a custom forecasting model — Compose layers, pick an optimizer, and run a training.

Pipelines

Train a pool of models on a schedule. Deploy the winner.

A pipeline binds a segment to a model pool and a schedule. Every run trains every model in the pool on the same data, scores them, and promotes the winner — automatically, or with your review.

Schedule on a calendar, on freshness, or on drift
Auto-promote winners, or require manual review before go-live
Every run is versioned — deploy any past training, not just the latest

Pipeline detail with a three-model pool, weekly schedule, latest training cycle scores, and history — Model pool, schedule, and the training history that produced the champion.

Multi-framework

TensorFlow, PyTorch, scikit, statsmodels, classic time-series — one workspace, mix freely.

Pretrained library

Forecasting, anomaly, classification, and embedding models — ready to fine-tune.

Adaptive cadence

Calendar, freshness, drift — pick a trigger and let the platform retrain itself.

Versioned everything

Trainings, fine-tunes, datasets, and configs remain addressable as versioned records.

Rollback-ready

Every retrain produces a clean rollback target. Promotion is reversible.

Reproducible training data

Every training points to a versioned segment, preserving the exact target, features, and time window.

Bring your own

Custom architectures live next to platform-provided ones — same workflow for both.

Eval-gated

Every training is evaluated against held-out data before it ever reaches production.

Put your data to work.

Connect your own data or start with a sample. Explore automatic insights, ask questions, engineer features, and build a forecast in the same workspace.

Start with your data Talk to sales