01
Typed nodes, enforced contracts
Data sources, inspections, transforms, splits, features, models, and evaluations are first-class building blocks. Invalid edges are blocked before they become confusing runs.
Forecasting workbench / self-hosted
Public release coming soon
Serious time series forecasting,
built on a canvas.
Compose end-to-end forecasting pipelines from typed nodes. Inspect data, split safely, engineer features, train statistical, ML, and deep learning models with prediction interval controls, compare results, and export artifacts - all running on infrastructure you own.
Powered by Polars
Demand Planning / monthly_pipeline
Run
Why teams pick TempusFlow
A workbench, not a notebook graveyard.
15
Core node types
3
Stat / ML / DL families
10
Inspection tabs
0
Lines of glue code required
The platform
A workbench that thinks like a forecaster.
Every node is typed. Every edge enforces a contract. The canvas blocks invalid wiring, and the executor applies leakage-safe ordering for split-aware transforms and feature engineering.
02
Three model families, one canvas
Statistical forecasting, sktime-powered ML, and Darts deep learning live in the same graph. Prediction interval controls are available across the model families, with clear warnings when a method cannot produce usable bounds.
03
Proper evaluation, not eyeballs
Single-holdout evaluation and walk-forward CV. SMAPE, MASE, RMSSE, mean error, residual diagnostics, PI coverage, Winkler score, horizon breakdowns, and directional accuracy.
04
Compare side-by-side
Train competing models on the same data with identical metrics. Focused versus full-history views help teams compare forecasts without losing temporal context.
05
Export. Reproduce. Govern.
Export readable Python or notebooks, save trained models to the registry, track versions and model cards, then run inference-only or retraining workflows from the same pipeline.
15 core node types
The full taxonomy. Nothing hidden.
Every primitive you need to take raw data to a versioned, evaluated, exported forecast. Drag any node onto the canvas, configure it, and keep the whole workflow visible.
Data Source
Files, demo data, Snowflake, DuckDB/MotherDuck, Postgres.
Data Inspection
10 tabs: quality, stats, shape, gaps, lags, outliers, ACF/PACF, seasonality, tests, decomp.
Transformation
13 transforms for cleaning, filtering, aggregating, resampling, filling, and differencing.
Split
By proportion or recent periods. Chronological.
Feature Engineering
Calendar, cyclic encoding, lags, rolling stats, holidays, interactions.
Statistical Model
AutoARIMA, Prophet, AutoETS, AutoTheta, TBATS, DOT, Random Walk, Drift, Seasonal Naive.
ML Model
Ridge, Huber, SVR, RF, GBM, HistGBM, XGBoost, LightGBM with tuning and coverage controls.
DL Model
RNN, LSTM, GRU, BlockRNN, TCN, N-HiTS, N-BEATS with Optuna and probabilistic options.
Backtesting
Walk-forward CV across multiple time windows.
Compare Baselines
Train selected models on the same data and rank them with identical metrics.
Compare Evaluations
Metric matrix, radar chart, and focused or full-history forecast overlay.
Reconciliation
Hierarchical coherence for panel and aggregate forecasts.
Ensemble
Combine multiple forecasts with average, median, trimmed mean, or accuracy weighting.
Evaluation
Forecasts, prediction intervals, residuals, metrics, tuning history, artifacts.
Output
Export models, forecasts, metrics, manifests, and run artifacts.
Self-hosted deployment
Your infrastructure.
Your forecasts. Your data.
Deploy the v1.0.0 stack on a customer-owned VM with Docker Compose. PostgreSQL, Redis, FastAPI, Celery, and S3-compatible storage stay under your control.
Application tierstateless
Web UI
React / TypeScript / React Flow
API Gateway
Node / Express / auth + RBAC
Forecasting Engine
Python / FastAPI / Celery
Polars Data Plane
columnar / typed / execution-ready
Modeling librariesopen-source
StatsForecast
AutoARIMA / AutoETS / AutoTheta
Darts
N-HiTS / N-BEATS / TCN / RNN
sktime
Reduction strategies / ML pipelines
Optuna
Hyperparameter tuning
Data tierstateful
PostgreSQL
metadata / registry / audit
Redis
queue / cache / session
S3-Compatible
artifacts / datasets / exports
Connectors
Snowflake / DuckDB / Postgres
Deploys on
docker compose up
PostgreSQL + Redis
S3-compatible artifacts
Hetzner demo track / customer single-VM track
Data engine
The pipeline engine is built around Polars.
TempusFlow uses Polars throughout the Python execution path for loading, transforming, inspecting, feature engineering, connector materialization, and artifact generation.
+
Columnar dataframes. Uploaded files, demo datasets, database query snapshots, transforms, and feature pipelines move through a fast columnar execution path.
+
Forecasting-aware transforms. Split handling, lag features, rolling windows, holiday features, and interactions are implemented with leakage-conscious Polars operations.
+
Parquet-first snapshots. Database query results and uploads are materialized into reproducible datasets before pipeline execution, so reruns know exactly what they trained on.
+
Clean handoff to model libraries. Polars data feeds StatsForecast, Prophet, sktime, and Darts through normalized contracts while preserving time, target, entity, and covariate metadata.
execution path / v1.0.0
From source to forecast artifacts
files / database snapshots / model results
Load and snapshot
CSV / Parquet / DB query
Transform and feature
split-safe Polars ops
Train and export
models / metrics / manifests
The goal is not a synthetic speed chart. It is a repeatable execution path: source data becomes a governed dataset, the graph runs in order, and artifacts are saved with lineage.
Enterprise
Built for governed forecasting teams.
RBAC, connection management, audit logs, a model registry, and a plugin system. The operational pieces are visible, configurable, and owned by the customer.
Full RBAC
Manage users, groups, project membership, owner/editor/viewer roles, and per-user capability flags from the admin console.
- project-scoped access
- node and capability controls
- admin-visible effective permissions
Model Registry
Save trained models, track versions, inspect scores and dataset labels, and run pipelines in inference-only or retraining mode.
- latest version badges
- archive and promote actions
- model cards with artifacts
Plugin System
Add your own node types. Two reference plugins ship with the platform: Hello World and a PyOD anomaly detector.
- typed node contracts
- manifest-driven UI
- custom result viewers
Audit Log
Track pipeline runs, model registry actions, connection changes, and permission events with actor, resource, timestamp, and details.
- admin filters
- CSV export
- resource-level history
Connections & Storage
Admins manage Snowflake, Postgres, and DuckDB/MotherDuck database connections, plus AWS S3 or S3-compatible storage for datasets and artifacts.
- connection test flow
- query preview and snapshots
- Azure Blob and GCS planned
Customer-Owned Deployment
Run TempusFlow on infrastructure your team controls, with data, artifacts, credentials, and operational policy kept inside your environment.
- customer-owned binaries
- offline-friendly operation
- no usage-based pricing
From raw data to evaluated forecast
Five steps. Same canvas.
STEP 01
Connect data
Start from files, demo datasets, or approved database queries from Snowflake, DuckDB/MotherDuck, and Postgres.
STEP 02
Inspect & transform
Review quality warnings, ACF/PACF, stationarity tests, seasonality detection, then split and engineer features.
STEP 03
Pick models
Choose statistical, ML, or DL models, or run several families on one canvas with tuned hyperparameters and PI settings.
STEP 04
Compare & backtest
Use walk-forward validation, residual diagnostics, side-by-side metrics, and comparison views to choose the strongest candidate.
STEP 05
Export & govern
Export Python or notebook code, register trained models, download forecasts and metrics, and preserve run artifacts.
TempusFlow v1.0.0 / self-hosted forecasting
Build forecast pipelines
your team can trust.
Public release coming soon. TempusFlow is being prepared as a customer-owned forecasting platform for teams that need evaluated, reproducible, governable forecasts without turning every workflow into bespoke notebook glue.