mt-data-mcp/docs/forecast/METHODS.md

Forecast methods reference

Audience: Operator

Catalog of --method keys for forecast_generate (and related tune/backtest tools): categories, libraries, default --params, and dependencies.

Plain-language method families: Theta · ARIMA · ETS · Monte Carlo / GBM · Foundation models · Ensemble.

Your install is the source of truth. Optional extras change what is available:

mtdata-cli forecast_list_methods --json
mtdata-cli forecast_list_methods --library statsforecast --json

The catalog includes unavailable methods by default and marks them with available=false plus the missing requirement. Pass --show-unavailable false for installed methods only. The default page is 20 methods; use --limit 100 when you need the complete catalog. Defaults below match the code; "auto" parameters are estimated from data unless you override them.

Related: Forecasting concepts · Parameter optimization · forecast_generate · Glossary

---

Method keys at a glance

Baselines/classical : naive, drift, seasonal_naive, theta, fourier_ols
Exp. smoothing/ARIMA: ses, holt, holt_winters_add, holt_winters_mul, ets, arima, sarima
Monte Carlo         : mc_gbm, hmm_mc
Analog / ensemble   : analog, ensemble
Machine learning    : mlforecast, mlf_rf, mlf_lightgbm
Foundation (pretr.) : chronos2, chronos_bolt, timesfm
Neural              : nhits, nbeatsx, tft, patchtst
StatsForecast       : statsforecast + sf_* family (sf_autoarima, sf_autoets, sf_theta, …)
Sktime              : sktime + skt_theta, skt_naive, skt_autoets

The sf_* and skt_* families are convenience aliases that pin a specific model within the StatsForecast / sktime libraries (e.g. sf_ets → AutoETS). Browse them with forecast_list_methods --library statsforecast.

---

Categories, libraries & dependencies

Category	Methods	Library	Optional dependency
Baseline / classical	naive, drift, seasonal_naive, theta, fourier_ols	native	none (core stack)
Exp. smoothing / ARIMA	ses, holt, holt_winters_add, holt_winters_mul, ets, arima, sarima	native	statsmodels
Monte Carlo	mc_gbm, hmm_mc	native	none (core stack)
Analog	analog	native	none (scipy/numpy/pandas)
Ensemble	ensemble	native	none
Machine learning	mlforecast, mlf_rf, mlf_lightgbm	mlforecast	mlforecast + scikit-learn / lightgbm
Foundation	chronos2, chronos_bolt, timesfm	pretrained	chronos / timesfm (+ torch)
Neural	nhits, nbeatsx, tft, patchtst	neuralforecast	neuralforecast (+ torch)
StatsForecast	statsforecast, sf_*	statsforecast	statsforecast
Sktime	sktime, skt_*	sktime	sktime

native methods ship with the lean core install. Everything else follows the dependency groups in SETUP.md / README. Use forecast_list_library_models --library <name> to enumerate the concrete models within a library-backed method.

---

Baseline & classical (native)

Fast, dependency-free references and solid intraday baselines.

Method	Key params (default)	Notes
naive	—	Last value carried forward
drift	—	Last value plus average drift
seasonal_naive	requires seasonality	Repeats the last seasonal cycle
theta	— (no user-settable method parameters)	Robust, fast general-purpose baseline; fitted alpha is reported in params_used
fourier_ols	terms=auto (≈min(3, m/2)), trend=true	OLS on Fourier terms; good for periodic series

mtdata-cli forecast_generate EURUSD --timeframe H1 --horizon 12 --method theta
mtdata-cli forecast_generate EURUSD --method fourier_ols --params "terms=4,trend=true"

---

Exponential smoothing & ARIMA (statsmodels)

Method	Key params (default)	Notes
ses	alpha=auto	Simple exponential smoothing
holt	damped=false; alpha,beta=auto	Linear-trend smoothing
holt_winters_add	damped=false; alpha,beta,gamma=auto; requires seasonality	Additive seasonality
holt_winters_mul	damped=false; alpha,beta,gamma=auto; requires seasonality	Multiplicative seasonality
ets	trend=add, seasonal=auto, damped=false; smoothing params auto	State-space ETS
arima	(p,d,q)=(1,1,1), trend=c when omitted	Non-seasonal ARIMA
sarima	(p,d,q)=(1,1,1); seasonal order defaults to (0,1,1,m) when seasonality > 1	Seasonal ARIMA

mtdata-cli forecast_generate EURUSD --method ets --params "trend=add,damped=true"

---

Monte Carlo simulation (native)

Simulation methods return distributional forecasts (paths + intervals), which pair naturally with the barrier tools.

Method	Key params (default)	Notes
mc_gbm	n_sims=500, seed=42, mu=auto, sigma=auto, ci_alpha=0.05	Geometric Brownian motion paths
hmm_mc	n_states=2, n_sims=500, seed=42, ci_alpha=0.05	Hidden-Markov regime-switched simulation

mtdata-cli forecast_generate EURUSD --method mc_gbm --params "n_sims=2000,seed=7"

Set seed explicitly for reproducible simulation runs.

---

Analog & ensemble (native)

Method	Key params (default)	Notes
analog	window_size=64, search_depth=5000, top_k=20, metric=euclidean, min_separation=window_size/4, ci_alpha=0.05	Nearest-neighbour analogs from history
ensemble	combines multiple methods	Blends member forecasts

mtdata-cli forecast_generate EURUSD --method analog --params "window_size=96,top_k=30"

---

Machine learning (mlforecast)

Gradient-boosted / tree models over lagged features. All MLForecast methods default their lags to 1..min(30, seasonality or 24) when lags is omitted.

Method	Key params (default)	Notes
mlforecast	model (required), lags=auto	Generic MLForecast wrapper; pick the underlying model
mlf_rf	n_estimators=200, max_depth=None, lags=auto	Random-forest regressor
mlf_lightgbm	n_estimators=200, learning_rate=0.05, num_leaves=31, max_depth=-1, lags=auto	LightGBM regressor

mtdata-cli forecast_generate EURUSD --method mlf_lightgbm --params "n_estimators=400,learning_rate=0.03"

---

Foundation models (pretrained)

Zero-shot pretrained forecasters. These download model weights on first use and benefit from a GPU (see ENV_VARS.md § Forecasting & GPU). Heavyweight fits are best run through the async training / model store.

Method	Key params (default)	Notes
chronos2	model_name=amazon/chronos-2	Amazon Chronos-2
chronos_bolt	model_name=amazon/chronos-bolt-base	Faster Chronos-Bolt
timesfm	model defaults	Google TimesFM (install via pip install -e .[forecast-timesfm])

mtdata-cli forecast_generate EURUSD --library pretrained --method chronos2 --horizon 24

---

Neural (neuralforecast)

Deep models via NeuralForecast. Not installed by requirements.txt or any extra today — install manually with pip install neuralforecast torch where the dependency stack resolves. On Windows Python 3.14 this currently fails because ray has no Windows cp314 wheels.

Method	Key params (default)	Notes
nhits	input_size=auto, max_steps=50, batch_size=32	N-HiTS
nbeatsx	input_size=auto, max_steps=50, batch_size=32	N-BEATSx
tft	input_size=auto, max_steps=50, batch_size=32	Temporal Fusion Transformer
patchtst	input_size=auto, max_steps=50, batch_size=32	PatchTST

---

StatsForecast & sktime families

Method	Key params (default)	Notes
statsforecast	model_name=autoarima, season_length=auto	Generic StatsForecast wrapper; pick the model
sf_*	fixed model per alias (e.g. sf_autoets, sf_theta, sf_garch)	season_length=auto
sktime	estimator=ThetaForecaster, estimator_params pass-through	Generic sktime wrapper
skt_theta / skt_naive / skt_autoets	fixed estimator per alias	Seasonality injected as sp when supported

mtdata-cli forecast_generate EURUSD --method sf_autoets --params "season_length=24"
mtdata-cli forecast_list_methods --library statsforecast --limit 100 --json  # full sf_* catalog

---

See Also

• FORECAST.md — Concepts, workflow, reproducibility, and libraries
• FORECAST.md § Parameter Optimization — forecast_tune_genetic, forecast_tune_optuna, forecast_optimize_hints
• BACKTESTING.md — Rolling backtests and metric selection
• UNCERTAINTY.md — Confidence and conformal intervals