A lightweight, batteries-included time-series forecasting toolkit for Python.

Unified model registry • Walk-forward backtesting • Probabilistic forecasting • CLI + Python API

Installation · Quick Start · Model Zoo · Architecture · Docs · Contributing

250+ models · 7 backends · 20+ metrics · 1 unified interface

Overview

• One interface across statistical, ML, and deep learning forecasters.
• Backtesting-first workflows with walk-forward CV, leaderboards, and intervals.
• First-class panel and global modeling with unique_id / ds / y long format.
• Minimal core install; heavier runtimes stay optional.
• Python API, CLI workflows, and reusable model artifacts share the same contracts.

---

Support Contract

ForeSight exposes a broad catalog, but the stable public surface is narrower than the full implementation footprint.

• Stable surface: the root foresight Python API, model registry metadata from foresight models list/info, the artifact schema/load contract, and the documented CLI workflows.
• CI-backed matrix: Python 3.10 and 3.11 for core installs and contract checks, plus packaging smoke on Python 3.10 and smoke coverage for stats and ml.
• Beta and experimental families remain available, but they are not covered by the same compatibility promise as the stable core workflows.

See Compatibility guide for the detailed support policy.

---

How It Works

---

Installation

pip install foresight-ts                # core (numpy + pandas only)

Optional extras:

Extra	Install	Use for
ml	pip install "foresight-ts[ml]"	scikit-learn models
xgb	pip install "foresight-ts[xgb]"	XGBoost models
lgbm	pip install "foresight-ts[lgbm]"	LightGBM models
catboost	pip install "foresight-ts[catboost]"	CatBoost models
stats	pip install "foresight-ts[stats]"	statsmodels models
torch	pip install "foresight-ts[torch]"	PyTorch models
sktime	pip install "foresight-ts[sktime]"	sktime adapter bridge
darts	pip install "foresight-ts[darts]"	Darts adapter bridge
gluonts	pip install "foresight-ts[gluonts]"	GluonTS adapter bridge
all	pip install "foresight-ts[all]"	everything above

Check the active environment, optional dependencies, and dataset resolution:

foresight doctor
foresight --data-dir /path/to/root doctor
foresight doctor --format text
foresight --data-dir /path/to/root doctor --format text --strict
foresight doctor --require-extra torch --strict

Install from source (for development)

git clone https://github.com/skygazer42/ForeSight.git
cd ForeSight
pip install -e ".[dev]"     # editable install + pytest, ruff, mypy

---

Quick Start

Python API

from foresight import eval_model, make_forecaster, make_forecaster_object

# Walk-forward evaluation on a built-in dataset
metrics = eval_model(
    model="theta", dataset="catfish", y_col="Total",
    horizon=3, step=3, min_train_size=12,
)
print(metrics)  # {'mae': ..., 'rmse': ..., 'mape': ..., 'smape': ...}

# Functional API — stateless forecaster
f = make_forecaster("holt", alpha=0.3, beta=0.1)
yhat = f([112, 118, 132, 129, 121, 135, 148, 148], horizon=3)

# Object API — fit / predict / save / load
obj = make_forecaster_object("moving-average", window=3)
obj.fit([1, 2, 3, 4, 5, 6])
yhat = obj.predict(3)

Beta Adapters

Beta integration adapters live under foresight.adapters rather than the stable root package surface.

from foresight.adapters import (
    make_sktime_forecaster_adapter,
    to_darts_timeseries,
    from_darts_timeseries,
    to_gluonts_list_dataset,
)

• make_sktime_forecaster_adapter(...) provides a minimal sktime-style fit/predict bridge.
• to_darts_timeseries(...) and from_darts_timeseries(...) convert between ForeSight data and Darts TimeSeries.
• to_gluonts_list_dataset(...) converts long-format history into a GluonTS ListDataset.

These adapters are intentionally beta and are not part of the root foresight.__all__ contract yet.

CLI

# Discover models and datasets
foresight models list                     # list all 250+ models
foresight models info theta               # inspect parameters & defaults
foresight models list --stability stable  # filter by maturity level
foresight models list --capability supports_x_cols=true --format json
foresight datasets list                   # browse built-in datasets
foresight doctor                          # inspect optional deps and data paths
foresight doctor --format text            # human-readable environment summary
foresight doctor --require-extra torch --strict

# Evaluate a model
foresight eval run --model theta --dataset catfish --y-col Total \
    --horizon 3 --step 3 --min-train-size 12

# Compare models on a leaderboard
foresight leaderboard models --dataset catfish --y-col Total \
    --horizon 3 --step 3 --min-train-size 12 \
    --models naive-last,seasonal-naive,theta,holt

# Forecast from any CSV
foresight forecast csv --model naive-last --path ./data.csv \
    --time-col ds --y-col y --parse-dates --horizon 7

# Cross-validation from any CSV
foresight cv csv --model naive-last --path ./data.csv \
    --time-col ds --y-col y --parse-dates \
    --horizon 3 --step-size 1 --min-train-size 24

# Detect anomalies from any CSV
foresight detect csv --path ./anomaly.csv \
    --time-col ds --y-col y --parse-dates \
    --score-method rolling-zscore --threshold-method zscore

Long-running CLI commands emit enhanced runtime logs to stderr by default, so stdout stays clean for JSON / CSV piping and file redirects.

More Python API examples (intervals, tuning, global models, hierarchical)

import pandas as pd
from foresight import (
    bootstrap_intervals,
    detect_anomalies,
    eval_hierarchical_forecast_df,
    forecast_model, tune_model, save_forecaster, load_forecaster,
    make_forecaster_object,
    make_global_forecaster, make_multivariate_forecaster,
    build_hierarchy_spec, reconcile_hierarchical_forecasts,
)

# Forecast with bootstrap prediction intervals
future_df = forecast_model(
    model="naive-last",
    y=[1, 2, 3, 4, 5, 6],
    ds=pd.date_range("2024-01-01", periods=6, freq="D"),
    horizon=3,
    interval_levels=(0.8, 0.9),
    interval_min_train_size=4,
)

# Save / load trained model artifacts
obj = make_forecaster_object("theta", alpha=0.3)
obj.fit([1, 2, 3, 4, 5])
save_forecaster(obj, "/tmp/theta.pkl")
loaded = load_forecaster("/tmp/theta.pkl")

# Detect anomalies from a dataset-backed series
anomalies = detect_anomalies(
    dataset="catfish",
    y_col="Total",
    model="naive-last",
    score_method="forecast-residual",
    min_train_size=24,
    step_size=1,
)

# Grid search tuning
result = tune_model(
    model="moving-average", dataset="catfish", y_col="Total",
    horizon=1, step=1, min_train_size=24, max_windows=8,
    search_space={"window": (1, 3, 6)},
)

# Multivariate model (VAR)
mv = make_multivariate_forecaster("var", maxlags=1)
yhat_mv = mv(wide_df[["sales", "traffic"]], horizon=2)

# Hierarchical reconciliation
hierarchy = build_hierarchy_spec(raw_df, id_cols=("region", "store"), root="total")
reconciled = reconcile_hierarchical_forecasts(
    forecast_df=pred_df, hierarchy=hierarchy,
    method="top_down", history_df=history_long,
)
hier_payload = eval_hierarchical_forecast_df(
    forecast_df=reconciled,
    hierarchy=hierarchy,
    y_col="y",
)

More CLI examples (artifacts, covariates, sweep)

# Forecast with prediction intervals
foresight forecast csv --model naive-last --path ./my.csv \
    --time-col ds --y-col y --parse-dates --horizon 3 \
    --interval-levels 80,90 --interval-min-train-size 12

# Enhanced runtime logs go to stderr; stdout remains pipe-safe
foresight forecast csv --model torch-mlp-direct --path ./train.csv \
    --time-col ds --y-col y --parse-dates --horizon 3 \
    --model-param lags=24 --model-param epochs=10 \
    --format json > /tmp/forecast.json

# Reduce per-epoch chatter or persist structured event logs
foresight eval run --model theta --dataset catfish --y-col Total \
    --horizon 3 --step 3 --min-train-size 12 \
    --no-progress --log-style plain --log-file /tmp/eval-log.jsonl

# Save and reuse model artifacts
foresight forecast csv --model naive-last --path ./my.csv \
    --time-col ds --y-col y --parse-dates --horizon 3 \
    --save-artifact /tmp/naive-last.pkl
foresight forecast artifact --artifact /tmp/naive-last.pkl --horizon 3

# Only load artifacts from trusted sources.

# Save and reuse a local x_cols artifact with the saved future covariates
foresight forecast csv --model sarimax --path ./my_exog.csv \
    --time-col ds --y-col y --parse-dates --horizon 3 \
    --model-param order=0,0,0 --model-param seasonal_order=0,0,0,0 \
    --model-param trend=c --model-param x_cols=promo \
    --save-artifact /tmp/sarimax.pkl
foresight forecast artifact --artifact /tmp/sarimax.pkl --horizon 2

# Override the saved future covariates with a new future CSV
foresight forecast artifact --artifact /tmp/sarimax.pkl \
    --future-path ./my_exog_future.csv --time-col ds --parse-dates \
    --horizon 4

# Reuse a quantile-capable global artifact and derive interval columns
foresight forecast artifact --artifact /tmp/xgb-global.pkl \
    --horizon 2 --interval-levels 80

# Override a saved global artifact with new future covariates
# The override CSV can contain canonical unique_id values or the raw id columns
# that were used when the artifact was saved, plus ds and required x_cols.
# Single-series global artifacts can also omit id columns entirely.
foresight forecast artifact --artifact /tmp/ridge-global.pkl \
    --future-path ./my_global_future.csv --time-col ds --parse-dates \
    --horizon 2

foresight artifact info --artifact /tmp/naive-last.pkl
foresight artifact info --artifact /tmp/naive-last.pkl --format markdown
foresight artifact validate --artifact /tmp/naive-last.pkl
foresight artifact diff \
    --left-artifact /tmp/naive-last.pkl \
    --right-artifact /tmp/naive-last-v2.pkl \
    --path-prefix metadata.train_schema.runtime --format csv
foresight artifact diff \
    --left-artifact /tmp/naive-last.pkl \
    --right-artifact /tmp/naive-last-v2.pkl \
    --path-prefix tracking_summary --format csv
foresight artifact diff \
    --left-artifact /tmp/naive-last.pkl \
    --right-artifact /tmp/naive-last-v2.pkl \
    --path-prefix future_override_schema --format markdown

# SARIMAX with exogenous features
foresight forecast csv --model sarimax --path ./my_exog.csv \
    --time-col ds --y-col y --parse-dates --horizon 3 \
    --model-param order=0,0,0 --model-param x_cols=promo

# Anomaly detection with exogenous covariates
foresight detect csv --model sarimax --path ./my_exog.csv \
    --time-col ds --y-col y --parse-dates \
    --score-method forecast-residual --threshold-method mad \
    --min-train-size 24 --model-param order=0,0,0 \
    --model-param seasonal_order=0,0,0,0 \
    --model-param trend=c --model-param x_cols=promo

# Multi-dataset sweep (parallel + resumable)
foresight leaderboard sweep \
    --datasets catfish,ice_cream_interest \
    --models naive-last,theta --horizon 3 --step 3 \
    --min-train-size 12 --jobs 4 --progress

# Conformal prediction intervals
foresight eval run --model theta --dataset catfish --y-col Total \
    --horizon 3 --step 3 --min-train-size 12 --conformal-levels 80,90

# Cross-validation predictions table
foresight cv run --model theta --dataset catfish --y-col Total \
    --horizon 3 --step-size 3 --min-train-size 12 --n-windows 30

# Cross-validation on arbitrary CSV data
foresight cv csv --model sarimax --path ./my_exog.csv \
    --time-col ds --y-col y --parse-dates \
    --horizon 3 --step-size 3 --min-train-size 24 \
    --model-param order=0,0,0 \
    --model-param seasonal_order=0,0,0,0 \
    --model-param trend=c --model-param x_cols=promo \
    --format json > /tmp/cv.json

---

Model Zoo

ForeSight organizes 250+ registered models into a few clear families. Use foresight models list and foresight models info <key> for the full catalog.

Model Landscape

Model Families

Family	Coverage
Statistical	naive baselines, moving averages, ETS/Holt-Winters, theta, Fourier, Kalman, intermittent demand
Machine learning	scikit-learn, XGBoost, LightGBM, CatBoost
Deep learning	MLP, RNN, CNN, Transformer, SSM, residual blocks
Panel/global	sklearn global, tree global, torch global
Multivariate / ensemble	VAR, mean, median, pipeline

Optional Models

scikit-learn — pip install "foresight-ts[ml]"

Local (lag-feature + direct multi-horizon):

ridge-lag, ridge-lag-direct, rf-lag, decision-tree-lag, extra-trees-lag, adaboost-lag, bagging-lag, lasso-lag, elasticnet-lag, knn-lag, gbrt-lag, hgb-lag, svr-lag, linear-svr-lag, kernel-ridge-lag, mlp-lag, huber-lag, quantile-lag, sgd-lag

Global/panel (step-lag, trains across all series):

ridge-step-lag-global, rf-step-lag-global, extra-trees-step-lag-global, gbrt-step-lag-global, decision-tree-step-lag-global, bagging-step-lag-global, lasso-step-lag-global, elasticnet-step-lag-global, knn-step-lag-global, svr-step-lag-global, mlp-step-lag-global, and more

XGBoost — pip install "foresight-ts[xgb]"

Strategy	Models
Direct	xgb-lag, xgb-dart-lag, xgbrf-lag, xgb-linear-lag
Recursive	xgb-lag-recursive, xgb-dart-lag-recursive, xgb-linear-lag-recursive
Step-index	xgb-step-lag
DirRec	xgb-dirrec-lag
MIMO	xgb-mimo-lag
Custom objectives	xgb-mae-lag, xgb-huber-lag, xgb-quantile-lag, xgb-poisson-lag, xgb-gamma-lag, ...
Global/panel	xgb-step-lag-global (supports quantile output)

LightGBM — pip install "foresight-ts[lgbm]"

Local: lgbm-lag, lgbm-lag-recursive, lgbm-step-lag, lgbm-dirrec-lag, lgbm-custom-lag

Global/panel: lgbm-step-lag-global (supports quantile output)

CatBoost — pip install "foresight-ts[catboost]"

Local: catboost-lag, catboost-lag-recursive, catboost-step-lag, catboost-dirrec-lag

Global/panel: catboost-step-lag-global (supports quantile output)

statsmodels — pip install "foresight-ts[stats]"

Family	Models
ARIMA	arima, auto-arima, sarimax, autoreg
Fourier Hybrid	fourier-arima, fourier-auto-arima, fourier-autoreg, fourier-sarimax
Decomposition	stl-arima, stl-autoreg, stl-ets, mstl-arima, mstl-autoreg, mstl-ets, mstl-auto-arima
TBATS-lite	tbats-lite, tbats-lite-autoreg, tbats-lite-auto-arima
Unobserved Components	uc-local-level, uc-local-linear-trend, uc-seasonal
ETS	ets
Multivariate	var

PyTorch — Local models — pip install "foresight-ts[torch]"

Category	Models
MLP / Linear	torch-mlp-direct, torch-nlinear-direct, torch-dlinear-direct, torch-tide-direct, torch-kan-direct
RNN	torch-lstm-direct, torch-gru-direct, torch-bilstm-direct, torch-bigru-direct, torch-attn-gru-direct
CNN	torch-cnn-direct, torch-tcn-direct, torch-resnet1d-direct, torch-wavenet-direct, torch-inception-direct, torch-scinet-direct
Transformer	torch-transformer-direct, torch-patchtst-direct, torch-crossformer-direct, torch-pyraformer-direct, torch-fnet-direct, torch-tsmixer-direct, torch-retnet-direct
Residual Blocks	torch-nbeats-direct, torch-nhits-direct
SSM / State-space	torch-mamba-direct, torch-rwkv-direct, torch-hyena-direct
Hybrid	torch-etsformer-direct, torch-esrnn-direct, torch-lstnet-direct
Probabilistic	torch-deepar-recursive, torch-qrnn-recursive
RNN Paper Zoo	100 named paper architectures (torch-rnnpaper-*-direct)
RNN Zoo	100 combos: 20 bases × 5 wrappers (torch-rnnzoo-*-direct)
Configurable Transformer	torch-xformer-*-direct — 10 attention variants × positional encodings × extras

PyTorch — Global/panel models — pip install "foresight-ts[torch]"

Train across all series in long-format; supports covariates (x_cols), time features, and optional quantile regression.

Category	Models
Transformer	torch-tft-global, torch-informer-global, torch-autoformer-global, torch-fedformer-global, torch-patchtst-global, torch-itransformer-global, torch-timesnet-global, torch-tsmixer-global
MLP / Linear	torch-nbeats-global, torch-nhits-global, torch-nlinear-global, torch-dlinear-global, torch-tide-global
RNN	torch-deepar-global, torch-lstnet-global, torch-esrnn-global
CNN	torch-tcn-global, torch-wavenet-global, torch-scinet-global
SSM	torch-mamba-global, torch-rwkv-global, torch-hyena-global

Multi-horizon Strategies

Strategy	How it works	Suffix
Direct	One model per horizon step	*-lag
Recursive	One-step model, iteratively re-fed	*-lag-recursive
Step-index	Single model with step as a feature	*-step-lag
DirRec	Per-step model with previous-step features	*-dirrec-lag
MIMO	Single model predicts entire horizon	*-mimo-lag

---

Architecture

---

Data Format

ForeSight uses a panel-friendly long format compatible with StatsForecast and Prophet:

Column	Description
unique_id	Series identifier (optional for single series)
ds	Timestamp (datetime)
y	Target value (float)
extra columns	Covariates / exogenous features

Covariate Roles

• historic_x_cols: observed only up to the forecast cutoff.
• future_x_cols: known through the forecast horizon.
• x_cols: shorthand alias for future_x_cols.

from foresight.data import to_long, prepare_long_df

long_df = to_long(
    raw_df, time_col="ds", y_col="y",
    id_cols=("store", "dept"),
    historic_x_cols=("promo_hist",),
    future_x_cols=("promo_plan", "price"),
)

prepared = prepare_long_df(
    long_df, freq="D",
    y_missing="interpolate",
    historic_x_missing="ffill",
    future_x_missing="ffill",
)

---

Evaluation & Backtesting

Capabilities

• Walk-forward backtesting with expanding or rolling windows.
• Full cross-validation tables with unique_id, ds, cutoff, step, y, yhat, model.
• Single-dataset leaderboards and multi-dataset sweeps.
• Grid-search tuning scored through backtesting.
• Conformal and bootstrap prediction intervals.

Metrics

Point Metrics	MAE · RMSE · MAPE · sMAPE · WAPE · MASE · RMSSE · MSE
Probabilistic Metrics	Pinball loss · CRPS · Coverage · Width · Sharpness · Interval Score · Winkler Score · Weighted Interval Score · MSIS

Probabilistic Forecasting

# Quantile regression with Torch global models
foresight eval run --model torch-itransformer-global \
    --dataset catfish --y-col Total --horizon 7 --step 7 --min-train-size 60 \
    --model-param quantiles=0.1,0.5,0.9

Produces yhat_p10, yhat_p50, yhat_p90 columns alongside yhat (defaults to median quantile).

---

Optional Dependencies

Extra	Backend	Version	Example Models
[ml]	scikit-learn	≥ 1.0	ridge-lag, rf-lag, hgb-lag, mlp-lag, *-step-lag-global
[xgb]	XGBoost	≥ 2.0	xgb-lag, xgb-step-lag, xgb-mimo-lag, xgb-step-lag-global
[lgbm]	LightGBM	≥ 4.0	lgbm-lag, lgbm-step-lag-global
[catboost]	CatBoost	≥ 1.2	catboost-lag, catboost-step-lag-global
[stats]	statsmodels	≥ 0.14	arima, auto-arima, sarimax, ets, var, stl-*, mstl-*
[torch]	PyTorch	≥ 2.0	torch-transformer-direct, torch-tft-global, torch-mamba-global
[all]	All of the above	—	—

---

Repository Structure

ForeSight/
├── src/foresight/              # Main Python package
│   ├── models/                 #   Model registry, catalog, factories
│   │   ├── catalog/            #     Model metadata shards
│   │   ├── factories.py        #     Runtime construction
│   │   └── registry.py         #     Public model facade
│   ├── services/               #   Forecast & evaluation orchestration
│   ├── contracts/              #   Validation & capability checks
│   ├── data/                   #   Long-format I/O & preprocessing
│   ├── datasets/               #   Built-in dataset registry
│   ├── features/               #   Feature engineering
│   ├── cli.py                  #   CLI entry point
│   ├── base.py                 #   BaseForecaster classes
│   ├── backtesting.py          #   Walk-forward engine
│   ├── metrics.py              #   20+ evaluation metrics
│   ├── conformal.py            #   Conformal prediction intervals
│   ├── hierarchical.py         #   Hierarchical reconciliation
│   └── transforms.py           #   Data transformations
├── examples/                   # Runnable example scripts
├── data/                       # Bundled CSV datasets
├── benchmarks/                 # Reproducible benchmark harness
├── tests/                      # 150+ test files
├── docs/                       # MkDocs documentation
└── tools/                      # Dev & release utilities

Example scripts: quickstart_eval.py · leaderboard.py · cv_and_conformal.py · intermittent_demand.py · torch_global_models.py · rnn_paper_zoo.py

---

Model Capability Flags

Model discovery surfaces machine-readable capability flags. Use them through foresight models list and foresight models info.

Capability keys

Flag	Meaning
supports_panel	Supports panel / multi-series workflows through the high-level long-format APIs
supports_univariate	Supports univariate forecasting targets
supports_multivariate	Supports multivariate / wide-matrix forecasting targets
supports_probabilistic	Supports probabilistic forecasting through intervals and/or quantiles
supports_conformal_eval	Can participate in conformal backtest evaluation flows
supports_future_covariates	Accepts known future covariates in supported workflows
supports_historic_covariates	Accepts historical covariate values in supported workflows
supports_static_covariates	Accepts series-level static covariates
supports_refit_free_cv	Can reuse training state across CV cutoffs without refitting from scratch
supports_x_cols	Accepts future covariates / exogenous regressors
supports_static_cols	Accepts series-level static covariates from long_df
supports_quantiles	Emits quantile forecast columns directly
supports_interval_forecast	Supports forecast intervals
supports_interval_forecast_with_x_cols	Supports forecast intervals when future covariates are provided
supports_artifact_save	Can be saved and reused through the artifact workflow
requires_future_covariates	Requires known future covariates

foresight models info holt-winters-add    # see all flags & defaults

---

Benchmarks

ForeSight ships a reproducible benchmark harness for packaged datasets:

python benchmarks/run_benchmarks.py --smoke          # CI smoke run
python benchmarks/run_benchmarks.py --config baseline --format md

---

Contributing

We welcome contributions! Here's how to get started:

git clone https://github.com/skygazer42/ForeSight.git
cd ForeSight
pip install -e ".[dev]"

ruff check src tests tools    # lint
ruff format src tests tools   # format
pytest -q                     # test

See docs/DEVELOPMENT.md for detailed guidelines and docs/ARCHITECTURE.md for the layered design.

---

Related Projects

ForeSight draws design inspiration from these excellent projects:

Project	Highlights
StatsForecast	Fast statistical baselines, unique_id/ds/y convention, cross-validation
NeuralForecast	Modern neural architectures (TFT, Informer, Autoformer, …)
Darts	Unified model API + backtesting helpers
sktime	Unified fit/predict interface and evaluation utilities
GluonTS	Probabilistic forecasting datasets + benchmarking
Prophet	ds/y DataFrame convention for forecasting
PyTorch Forecasting	Deep learning forecasting pipelines
Kats	Time series analysis & forecasting toolbox (Meta)

---

Citing ForeSight

If you find ForeSight useful in your research, please consider citing:

@software{foresight2024,
  title   = {ForeSight: A Lightweight Time-Series Forecasting Toolkit},
  author  = {ForeSight Contributors},
  year    = {2024},
  url     = {https://github.com/skygazer42/ForeSight},
  version = {0.2.9},
}

---

License: GPL-3.0-only

Issues · Documentation