milt - Modern Integrated Library for Timeseries

milt is a unified R framework for time series forecasting, anomaly detection, classification, clustering, diagnostics, and deployment. It is designed to give R users a single, consistent interface across classical statistical models, machine learning backends, and deep-learning integrations.

The package follows one core idea: users should not need to learn a different API for each model family. Whether you are fitting a naive baseline, an Auto-ARIMA model, an XGBoost regressor, or an ensemble, the workflow stays consistent.

Documentation site: https://ntigideon.github.io/milt

What milt covers

milt is intended to cover the end-to-end time series workflow:

• series creation and conversion from ts, tibble, data.frame, tsibble, xts, zoo, and numeric vectors
• forecasting with classical, ML, and deep-learning backends
• train/test splitting, rolling-origin backtesting, and model comparison
• feature engineering with lag, rolling, Fourier, calendar, and scaling steps
• anomaly detection and changepoint analysis
• multi-series clustering, classification, and hierarchical reconciliation
• diagnostics, exploratory analysis, and model explainability
• saving, loading, dashboarding, and serving models

Installation

# install.packages("pak")
pak::pak("ntiGideon/milt")

milt keeps heavy backend dependencies optional so the core package installs quickly and reliably. Use milt_install_backends() to install what you need:

library(milt)

milt_install_backends()                        # everything
milt_install_backends("forecasting")           # forecast + prophet
milt_install_backends("ml")                    # xgboost, lightgbm, glmnet, ranger, e1071
milt_install_backends("deep_learning")         # torch  (then run torch::install_torch())
milt_install_backends("extras")                # anomaly, clustering, causal impact
milt_install_backends("reporting")             # rmarkdown, shiny, plumber
milt_install_backends("prophet")               # a single package by name

Group	Packages
"forecasting"	forecast, prophet
"ml"	xgboost, lightgbm, glmnet, ranger, e1071
"deep_learning"	torch
"extras"	isotree, dbscan, dtw, CausalImpact, changepoint
"reporting"	rmarkdown, shiny, plumber, jsonlite

Core workflow

The central workflow is:

library(milt)

series <- milt_series(AirPassengers)

fct <- milt_model("auto_arima") |>
  milt_fit(series) |>
  milt_forecast(horizon = 12)

print(fct)
plot(fct)

The same shape is used throughout the package:

milt_model("<name>") |> milt_fit(series) |> milt_forecast(horizon = 12)

Main objects

The package is organized around a small set of core objects:

• MiltSeries: the canonical time series container
• MiltModel: a fitted or unfitted model object
• MiltForecast: forecast results with point forecasts and intervals
• MiltBacktest: rolling-origin validation results
• MiltComparison: ranked comparison across multiple models
• MiltAnomalies: anomaly detection output

This design keeps downstream workflows stable. The same plotting, printing, and coercion patterns apply across most outputs.

Quick start

library(milt)

# 1. Create a series
air <- milt_series(AirPassengers)

# 2. Inspect and diagnose
print(air)
dx <- milt_diagnose(air)
print(dx)

# 3. Fit and forecast
fct <- milt_model("ets") |>
  milt_fit(air) |>
  milt_forecast(horizon = 24)

print(fct)
plot(fct)

# 4. Evaluate on a split
spl <- milt_split(air, ratio = 0.8)
fct_test <- milt_model("naive") |>
  milt_fit(spl$train) |>
  milt_forecast(spl$test$n_timesteps())

milt_accuracy(
  actual = spl$test$values(),
  predicted = fct_test$as_tibble()$.mean
)

# 5. Backtest a model
bt <- milt_backtest(
  model = milt_model("naive"),
  series = air,
  horizon = 12,
  initial_window = 108L,
  stride = 12L
)

print(bt)
plot(bt)

Feature engineering and pipelines

milt supports both direct feature steps and composed pipelines.

Direct feature engineering:

air_features <- air |>
  milt_step_lag(lags = 1:12) |>
  milt_step_rolling(windows = c(3L, 6L), fns = c("mean", "sd")) |>
  milt_step_fourier(period = 12, K = 2)

Pipeline workflow:

pipe <- milt_pipeline() |>
  milt_pipe_step_lag(lags = 1:12) |>
  milt_pipe_step_rolling(windows = 3L, fns = "mean") |>
  milt_pipe_model("xgboost")

Model families

The package includes or supports:

• baseline models such as naive, snaive, and drift
• classical statistical models such as ets, auto_arima, theta, tbats, and stl
• machine-learning backends such as xgboost, lightgbm, ranger, glmnet, svm, and KNN-based methods
• ensemble workflows
• deep-learning integrations through torch and reticulate-based backends

Use list_milt_models() to inspect the currently registered model set.

Diagnostics, anomaly detection, and explainability

The package is not limited to forecasting. It also includes:

• milt_diagnose() for stationarity, trend, seasonality, gaps, and outlier summaries
• milt_detector() and milt_detect() for anomaly detection
• milt_changepoints() for changepoint analysis
• milt_eda() for exploratory analysis
• milt_explain() for model explainability and feature importance

Multi-series workflows

For grouped or panel time series, milt supports:

• grouped MiltSeries construction
• clustering with milt_cluster()
• classification with milt_classifier(), milt_classify_fit(), and milt_classify_predict()
• reconciliation with milt_reconcile()

Deployment and reporting

Trained objects can be operationalized through:

• milt_save() and milt_load()
• milt_serve() for API-style serving
• milt_dashboard() for interactive exploration
• milt_report() for report generation

Built-in datasets

data(milt_air)
data(milt_retail)
data(milt_energy)

• milt_air: monthly airline passenger counts
• milt_retail: synthetic grouped retail sales data
• milt_energy: synthetic hourly energy demand with related covariates

Documentation map

The website is organized around both reference pages and longer-form articles.

Start with:

• Getting started
• Forecasting
• Pipelines
• Multi-series workflows
• Anomaly detection
• Deep learning
• Deployment

The reference index documents the exported functions, result objects, metrics, datasets, and S3 methods.

Architecture

At a high level:

Layer 1  User API
         milt_series(), milt_model(), milt_fit(), milt_forecast(), ...

Layer 2  Core objects
         MiltSeries, MiltModel, MiltForecast, MiltBacktest, MiltComparison

Layer 3  Feature and workflow engine
         lag/rolling/Fourier/calendar/scale steps, pipelines, backtesting

Layer 4  Backends
         classical statistical models, ML backends, deep-learning integrations

Layer 5  Delivery
         metrics, explainability, anomaly detection, saving, serving, reporting

This separation is what allows the package to expose one user-facing interface while still supporting many implementation backends.

Contributing

Bug reports and pull requests are welcome at https://github.com/ntiGideon/milt/issues.

Please review:

• CONTRIBUTING.md
• CODE_OF_CONDUCT.md

License

MIT License.