AutoTSForecast

Automated Time Series Forecasting with Per-Series Model Selection

AutoTSForecast automatically finds the best forecasting model for each of your time series. No more guessing whether Prophet, ARIMA, XGBoost, or Chronos-2 foundation model works best — let the algorithm decide. New: Zero-shot forecasting with Chronos-2 — no training required, just pass your data and get state-of-the-art predictions!

🚀 Key Features

Feature	Description	Benefit
MCP Server 🆕	Plug into Claude Desktop, Cursor, Windsurf via Model Context Protocol	Any AI agent forecasts your data natively
OpenAI / Anthropic Tools 🆕	Ready-made function-calling schemas	GPT & Claude can call forecasting tools directly
LangChain Integration 🆕	BaseTool wrappers for any LangChain agent	Build agentic pipelines in minutes
FastAPI REST Service 🆕	HTTP endpoints for every forecasting operation	Language-agnostic agent integration
Anomaly Detection 🆕	Z-score, IQR, Isolation Forest, forecast-residual	Clean data before forecasting
NLP Insight Engine 🆕	Plain-English forecast summaries (rule-based + LLM)	Agents explain forecasts in natural language
Model Registry 🆕	Save, load, list, and delete fitted models locally	Fit once, reuse anywhere
Structured Outputs 🆕	Pydantic models for all results	Machine-readable, JSON-serialisable, agent-ready
Chronos-2 Foundation Model	Zero-shot forecasting with pre-trained models (9M-710M params)	No training needed — just pass your data!
Per-Series Model Selection	Automatically pick the best model for each series	Different series, different patterns → optimal accuracy
Per-Series Covariates	Pass different features to different series	Products driven by different factors get custom features
Prediction Intervals	Conformal prediction with coverage guarantees	Quantify uncertainty without assumptions
Calendar Features	Auto-extract day-of-week, month, holidays	Handle seasonality automatically
Hierarchical Reconciliation	Ensure forecasts add up (total = sum of parts)	Coherent forecasts across organizational levels
Parallel Processing	Fit many series simultaneously	Scale to thousands of series
Interpretability	Sensitivity analysis & SHAP	Understand what drives your forecasts

✨ What's New in v0.5.0 — Agentic AI Edition

• 🤖 MCP Server — autotsforecast-mcp CLI connects directly to Claude Desktop, Cursor, and Windsurf. 7 tools: fit & forecast, backtest, prediction intervals, anomaly detection, calendar features, hierarchy reconciliation, model catalog.
• 🔧 OpenAI & Anthropic Tool Schemas — Drop-in get_openai_tools() / get_anthropic_tools() for GPT and Claude. handle_tool_call() dispatcher handles everything.
• 🦜 LangChain Tools — get_autotsforecast_tools() returns BaseTool instances for any LangChain ReAct or LCEL agent.
• 🌐 FastAPI REST Service — autotsforecast-api CLI starts an HTTP server at any host/port. 8 endpoints covering every operation.
• 📡 Anomaly Detection — AnomalyDetector with four methods. Detects outliers before forecasting to protect model accuracy.
• 💬 InsightEngine — Rule-based trend/risk analysis + optional LLM narrative for any LLM client.
• 📦 ModelRegistry — registry.save(auto, name="v1") / registry.load("v1") for local model persistence with JSON index.
• 📐 Structured Outputs — auto.to_structured() returns a Pydantic ForecastResult. All new APIs are JSON-serialisable.

✨ What's New in v0.4.0

• 📓 Rewritten tutorial — examples/autotsforecast_tutorial.ipynb redesigned with a DGP that guarantees measurable improvements for per-series covariates and hierarchical reconciliation
• 📦 Portable notebook — Added pip install autotsforecast[ml] installation cell so the notebook runs anywhere without this repo
• 📚 Docs overhaul — All documentation files updated: corrected model tables, covariate support flags, Chronos-2 details
• 🐛 Bug fixes — get_summary() / print_summary() now work correctly in per-series mode
• 🐛 Bug fixes — BacktestValidator now clones the model per fold (no shared-state mutation)
• 🐛 Bug fixes — VARForecaster raises a clear error when fewer than 2 series are provided
• ⚙️ Internals — Version sourced from package metadata (single source of truth)
• 🔧 CI/CD — GitHub Actions workflow runs the full test suite on every push/PR

✨ What's New in v0.3.8+

• 🚀 Chronos-2 Foundation Model — Zero-shot forecasting with state-of-the-art pre-trained models (no training needed!)
• 🎯 Per-Series Covariates — Pass different features to different series via X={series: df}
• 📊 Prediction Intervals — Conformal prediction for uncertainty quantification
• 📅 Calendar Features — Automatic time-based feature extraction with cyclical encoding
• 🖼️ Better Visualization — Static (matplotlib) and interactive (Plotly) forecast plots
• ⚡ Parallel Processing — Speed up multi-series forecasting with joblib
• 📈 Progress Tracking — Rich progress bars for long-running operations

Installation

🚀 Recommended: Install Everything

pip install "autotsforecast[all]"

This installs all 10 models plus visualization, interpretability, and new features.

🤖 Agentic AI Features (v0.5.0)

# MCP server — connect to Claude Desktop, Cursor, Windsurf
pip install "autotsforecast[mcp]"

# FastAPI REST service — HTTP interface for any agent or app
pip install "autotsforecast[api]"

# LangChain tools — for LangChain / LCEL agents
pip install "autotsforecast[langchain]"

# All agentic integrations in one shot
pip install "autotsforecast[agentic]"

# Streamlit web app (no-code UI)
pip install "autotsforecast[app]"

🖥️ Streamlit Web App

autotsforecast ships with a full no-code web UI built with Streamlit. It is not imported as a Python module — you run it as a web server:

pip install "autotsforecast[app]"
git clone https://github.com/weibinxu86/autotsforecast
cd autotsforecast
python3 -m streamlit run streamlit_app.py
# → opens http://localhost:8501

What the app includes:

• Upload any CSV or use built-in demo data
• Select target columns and (optionally) per-series covariates
• Choose from 9 model types with a dropdown
• Backtest toggle + per-series best-model table
• What-if scenario comparison (up to 5 scenarios)
• Download forecast + metrics as CSV

For a minimal 80-line example you can customise, see my_minimal_app.py (generated by examples/agentic_tutorial.ipynb Step 9 — run the notebook cell, then cd autotsforecast && python3 -m streamlit run my_minimal_app.py).

pip install autotsforecast

This gives you 6 models out of the box:

Model	Description
ARIMAForecaster	Classical ARIMA
ETSForecaster	Exponential smoothing
LinearForecaster	Linear regression — requires covariates X
MovingAverageForecaster	Simple baseline
RandomForestForecaster	ML with covariates ✓
VARForecaster	Vector autoregression — requires ≥ 2 series

Install Specific Optional Models

Some models require additional dependencies:

# Add XGBoost (gradient boosting with covariates)
pip install "autotsforecast[ml]"

# Add Prophet (Facebook's forecasting library)
pip install "autotsforecast[prophet]"

# Add LSTM (deep learning)
pip install "autotsforecast[neural]"

# Add Chronos-2 (foundation model - state-of-the-art zero-shot forecasting)
pip install "autotsforecast[chronos]"

# Add SHAP (interpretability)
pip install "autotsforecast[interpret]"

# Add visualization tools (Plotly, progress bars)
pip install "autotsforecast[viz]"

Model Availability Summary

Model	Basic Install	Extra Required
ARIMA, ETS, Linear*, MovingAverage, RandomForest, VAR	✅	—

* LinearForecaster requires covariates X to be passed (it is not included in get_default_candidate_models()). | XGBoostForecaster | ❌ | pip install "autotsforecast[ml]" | | ProphetForecaster | ❌ | pip install "autotsforecast[prophet]" | | LSTMForecaster | ❌ | pip install "autotsforecast[neural]" | | Chronos2Forecaster | ❌ | pip install "autotsforecast[chronos]" | | SHAP Analysis | ❌ | pip install "autotsforecast[interpret]" | | Interactive Plots | ❌ | pip install "autotsforecast[viz]" | | MCP Server | ❌ | pip install "autotsforecast[mcp]" | | FastAPI REST | ❌ | pip install "autotsforecast[api]" | | LangChain Tools | ❌ | pip install "autotsforecast[langchain]" |

Quick Start

1. AutoForecaster — Let the Algorithm Choose

from autotsforecast import AutoForecaster
from autotsforecast.models.base import MovingAverageForecaster
from autotsforecast.models.external import ARIMAForecaster, ProphetForecaster, RandomForestForecaster, Chronos2Forecaster

# Your time series data (pandas DataFrame)
# y = pd.DataFrame({'series_a': [...], 'series_b': [...]})

# Define candidate models (including Chronos-2 foundation model)
candidates = [
    ARIMAForecaster(horizon=14),
    ProphetForecaster(horizon=14),
    RandomForestForecaster(horizon=14, n_lags=7),
    MovingAverageForecaster(horizon=14, window=7),
    Chronos2Forecaster(horizon=14, model_name='autogluon/chronos-2-small'),  # Zero-shot foundation model
]

# AutoForecaster picks the best model across all series (default)
auto = AutoForecaster(candidate_models=candidates, metric='rmse')
auto.fit(y_train)
forecasts = auto.forecast()

# See which model was selected
print(auto.best_model_name_)  # e.g., 'Chronos2Forecaster'

# OR: Pick the best model for EACH series separately
auto = AutoForecaster(candidate_models=candidates, metric='rmse', per_series_models=True)
auto.fit(y_train)
forecasts = auto.forecast()

# See which models were selected per series
print(auto.best_model_names_)  # e.g., {'series_a': 'Chronos2Forecaster', 'series_b': 'ARIMAForecaster'}

2. Using Covariates (External Features)

from autotsforecast.models.external import XGBoostForecaster

# X contains external features (temperature, promotions, etc.)
model = XGBoostForecaster(horizon=14, n_lags=7)
model.fit(y_train, X=X_train)
forecasts = model.predict(X=X_test)

Models supporting covariates: Prophet, XGBoost, RandomForest, Linear

2.1 Calendar Features

Automatic time-based feature extraction:

from autotsforecast.features.calendar import CalendarFeatures

# Auto-detect features with cyclical encoding
cal = CalendarFeatures(cyclical_encoding=True)
features = cal.fit_transform(y_train)

# Generate future features for forecasting
future_features = cal.transform_future(horizon=30)

2.2 Per-Series Covariates — Different Features for Each Series

Use Case: When different time series are driven by different external factors.

from autotsforecast import AutoForecaster
from autotsforecast.models.base import MovingAverageForecaster
from autotsforecast.models.external import RandomForestForecaster, XGBoostForecaster

# Example: Forecasting sales for different products
# Product A: Summer product (driven by weather and advertising)
X_product_a = pd.DataFrame({
    'temperature': [...],      # Weather matters for Product A
    'advertising_spend': [...] # Marketing campaigns
}, index=dates)

# Product B: Everyday product (driven by pricing and promotions)
X_product_b = pd.DataFrame({
    'competitor_price': [...],  # Price competition matters for Product B
    'promotion_active': [...]   # Promotional events
}, index=dates)

# Create dictionary mapping each series to its covariates
X_train_dict = {
    'product_a_sales': X_product_a_train,
    'product_b_sales': X_product_b_train
}

X_test_dict = {
    'product_a_sales': X_product_a_test,
    'product_b_sales': X_product_b_test
}

# Define candidate models (all support covariates X)
candidates = [
    RandomForestForecaster(horizon=14, n_lags=7),
    XGBoostForecaster(horizon=14, n_lags=7),
    MovingAverageForecaster(horizon=14, window=7),  # covariate-free baseline
]

# AutoForecaster with per-series model selection
auto = AutoForecaster(
    candidate_models=candidates,
    per_series_models=True,  # Select best model for each series
    metric='rmse'
)

# Fit: Each series uses its own covariates
auto.fit(y_train, X=X_train_dict)

# Forecast: Provide future covariates for each series
forecasts = auto.forecast(X=X_test_dict)

# See which model was selected for each series
print(auto.best_model_names_)
# Output: {'product_a_sales': 'RandomForestForecaster', 
#          'product_b_sales': 'XGBoostForecaster'}

Key Benefits:

• ✅ Each series uses only relevant features (reduces noise)
• ✅ Better accuracy through targeted feature engineering
• ✅ Handle heterogeneous products with different drivers
• ✅ Scalable to large portfolios with diverse characteristics
• ✅ Backward compatible: still works with single DataFrame for all series

3. Hierarchical Reconciliation

Ensure forecasts add up correctly (e.g., total = region_a + region_b):

from autotsforecast.hierarchical.reconciliation import HierarchicalReconciler

hierarchy = {'total': ['region_a', 'region_b']}
reconciler = HierarchicalReconciler(forecasts=base_forecasts, hierarchy=hierarchy)
reconciler.reconcile(method='ols')
coherent_forecasts = reconciler.reconciled_forecasts

4. Backtesting (Cross-Validation)

from autotsforecast.backtesting.validator import BacktestValidator

validator = BacktestValidator(model=my_model, n_splits=5, test_size=14)
validator.run(y_train, X=X_train)

# Get results
results = validator.get_fold_results()  # RMSE per fold
print(f"Average RMSE: {results['rmse'].mean():.2f}")

5. Interpretability (Feature Importance)

from autotsforecast.interpretability.drivers import DriverAnalyzer

analyzer = DriverAnalyzer(model=fitted_model, feature_names=['temperature', 'promotion'])
importance = analyzer.calculate_feature_importance(X_test, y_test, method='sensitivity')

6. Prediction Intervals

Generate prediction intervals with conformal prediction:

from autotsforecast.uncertainty.intervals import PredictionIntervals

# After fitting a model
pi = PredictionIntervals(method='conformal', coverage=[0.80, 0.95])
pi.fit(model, y_train)
intervals = pi.predict(forecasts)

# Access intervals
print(intervals['lower_95'], intervals['upper_95'])

7. Chronos-2 Foundation Model (Zero-Shot Forecasting)

State-of-the-art pretrained model - no training needed!

from autotsforecast.models.external import Chronos2Forecaster

# Initialize with default model (120M params, best accuracy)
model = Chronos2Forecaster(
    horizon=30,
    model_name="amazon/chronos-2"  # or "autogluon/chronos-2-small" for faster inference
)

# Fit (just stores context, no training!)
model.fit(y_train)

# Generate point forecasts (median)
forecasts = model.predict()

# Generate probabilistic forecasts with uncertainty quantification
quantile_forecasts = model.predict_quantiles(quantile_levels=[0.1, 0.5, 0.9])
# Returns: value_q10, value_q50, value_q90 columns

Available Model Sizes:

• amazon/chronos-2 - 120M params (best accuracy)
• autogluon/chronos-2-small - 28M params (balanced, tested: 0.63% MAPE)
• amazon/chronos-bolt-tiny - 9M params (ultra fast)
• amazon/chronos-bolt-small - 48M params (balanced speed/accuracy)
• amazon/chronos-bolt-base - 205M params (high accuracy + fast)

Why Chronos-2?

• ✅ Zero-shot: No training required
• ✅ State-of-the-art accuracy on multiple benchmarks
• ✅ Built-in uncertainty quantification
• ✅ Multiple model sizes for different use cases

8. Visualization

Create publication-ready plots:

from autotsforecast.visualization.plots import plot_forecast, plot_forecast_interactive

# Static matplotlib plot
fig = plot_forecast(y_train, y_test, forecast, lower=lower_95, upper=upper_95)

# Interactive Plotly plot
fig = plot_forecast_interactive(y_train, y_test, forecast)
fig.show()

9. Parallel Processing

Speed up multi-series forecasting:

from autotsforecast.utils.parallel import ParallelForecaster, parallel_map

# Create parallel forecaster
pf = ParallelForecaster(n_jobs=4)

# Fit each series in parallel
fitted_models = pf.parallel_series_fit(
    model_factory=lambda: RandomForestForecaster(horizon=14),
    y=y_train,
    X=X_train
)

🤖 Agentic AI — v0.5.0

Anomaly Detection

Clean your data before forecasting:

from autotsforecast.anomaly.detector import AnomalyDetector

detector = AnomalyDetector(method='zscore', contamination=0.05)
anomalies = detector.fit_predict(y_train)  # bool DataFrame
summary = detector.get_summary()           # AnomalyResult (Pydantic)
print(f"Found {summary.total_anomalies} anomalies")

Structured Outputs

Get machine-readable results from AutoForecaster:

auto = AutoForecaster(candidates, metric='rmse')
auto.fit(y_train)
forecasts = auto.forecast()
result = auto.to_structured()   # ForecastResult (Pydantic)
print(result.model_dump_json()) # Perfect for agents / REST APIs

Natural Language Insights

from autotsforecast.nlp.insights import InsightEngine

engine = InsightEngine(mode='rule_based')
summary = engine.summarize_forecast_dataframes(y_train, forecasts, y_test)
risks   = engine.flag_risks_from_dataframes(y_train, forecasts)

Model Registry

Save and reload fitted models:

from autotsforecast.registry.store import ModelRegistry

registry = ModelRegistry()
registry.save(auto, name='production_v1', tags={'version': '1.0'})

# Later, in a different process or deployment:
auto_loaded = registry.load('production_v1')
new_forecasts = auto_loaded.forecast()

MCP Server (Claude Desktop / Cursor / Windsurf)

# Install
pip install "autotsforecast[mcp]"

# Start server (stdio transport)
autotsforecast-mcp

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "autotsforecast": {
      "command": "autotsforecast-mcp"
    }
  }
}

Claude can then use 7 tools: fit_and_forecast, run_backtest, prediction_intervals, anomaly_detection, calendar_features, reconcile_hierarchy, model_catalog.

FastAPI REST Service

# Install
pip install "autotsforecast[api]"

# Start server (default: http://0.0.0.0:8000)
autotsforecast-api

Endpoints: GET /health, GET /models, POST /forecast, POST /backtest, POST /intervals, POST /anomalies, POST /calendar-features, POST /reconcile.

OpenAI / Anthropic Tool Calling

from autotsforecast.integrations.openai_schemas import (
    get_openai_tools, get_anthropic_tools, handle_tool_call
)

# OpenAI
tools = get_openai_tools()
# response = openai.chat.completions.create(model="gpt-4o", tools=tools, ...)
# result = handle_tool_call(tool_name, arguments)

# Anthropic
tools = get_anthropic_tools()
# response = anthropic.messages.create(tools=tools, ...)

LangChain Integration

from autotsforecast.integrations.langchain_tools import get_autotsforecast_tools

tools = get_autotsforecast_tools()
# Pass to any LangChain ReAct or LCEL agent
# agent = create_react_agent(llm, tools, prompt)

Requirements

• Python ≥ 3.8
• Core: numpy, pandas, scikit-learn, statsmodels, scipy, joblib

License

MIT License

Contributing

Contributions welcome! Visit the GitHub repository to get started.

@software{autotsforecast2026,
  title={AutoTSForecast: Automated Time Series Forecasting},
  author={Weibin Xu},
  year={2026},
  url={https://github.com/weibinxu86/autotsforecast}
}