Modeltime Ecosystem Roadmap: New Algorithms & Models

[mdancho84]

Modeltime Ecosystem Roadmap

The modeltime project roadmap tracks the overall development of the Modeltime Ecosystem of forecasting packages. Modeltime is a cutting-edge ecosystem for forecasting using strategies and best practices that won or placed highly in major forecasting competitions. We have a state-of-the-art Time Series Forecasting Course (DS4B 203-R) that teaches Machine Learning, Deep Learning, and Feature Engineering for Time Series. Take this course to become the forecasting expert for your organization.

Forecasting Approaches

• [x] Global Models: Panel data and global modeling was introduced in Modeltime >= 0.3.0 and Modeltime Ensemble >= 0.3.0. We now have a vignette covering Forecasting with Panel Data.
• [x] Iterative Forecasting: The sknifedatar package extends modeltime for iterative forecasting. We are also working on an iterative forecasting solution that can be tracked with Issue #122. Nested Forecasting has been implemented. Refer to the Nested Forecasting Tutorial.

Ensembles (Stacking, Averaging)

Please refer to modeltime.ensemble R package: https://business-science.github.io/modeltime.ensemble/

• [x] Stacking - See modeltime.ensemble::ensemble_model_spec()
• [x] Averaging - See modeltime.ensemble::ensemble_average()
• [x] Weighted Averaging - See modeltime.ensemble::ensemble_weighted()

Forecast

These algorithms are provided in the base modeltime package. Refer to included models in the model list: https://business-science.github.io/modeltime/articles/modeltime-model-list.html

• [x] STLM - Decomposition Models - See seasonal_reg()
• [x] TBATS - See seasonal_reg()
• [x] NNETAR - See nnetar_reg()
• [x] Theta Method - See exp_smoothing() (#93)
• [x] CROSTON - See exp_smoothing() (#98)

GluonTS (Deep Learning)

Please refer to modeltime.gluonts R package: https://github.com/business-science/modeltime.gluonts

• [x] DeepAR - See modeltime.gluonts::deep_ar()
• [x] N-BEATS - See modeltime.gluonts::nbeats()

Autoregressive (Recursive) Forecasting

Recursive forecasting is required when Lags of Features < Forecast Horizon. Please refer to the recursive() function that allows any model to be converted into a recursive prediction model: https://business-science.github.io/modeltime/reference/recursive.html

• [x] Recursive for Single Time Series - Can be implemented using recursive()
• [x] Recursive for Panel Data (Multiple Time Series) - Follow #67
• [x] Recursive Ensembles - Tracked with #86. Refer to modeltime.ensemble for development efforts. https://business-science.github.io/modeltime.ensemble/

Resampling & Backtesting

• [x] Please refer to modeltime.resample for backtesting and resampling time series: https://business-science.github.io/modeltime.resample/index.html

Hyper Parameter Tuning & Parallel Processing

• [x] We have new hyperparameter tuning & parallel processing functionality that is available in Modeltime >= 0.6.0.
• [x] Vignette: Hyperparameter Tuning and Parallel Processing

Global Baseline Models

Refer to baseline algorithms #37

• [x] Naive & Seasonal Naive - See naive_reg()
• [x] Mean & Median Window Forecasts - See window_reg()

H2O

This project is being managed via modeltime.h2o: https://github.com/business-science/modeltime.h2o

• [x] H2O AutoML Backend (business-science/modeltime.h2o#1)

General Additive Models (GAMs)

• [x] Parsnip now incorporates GAMs

GARCH Models

Refer to garchmodels (Website, GitHub).

• [x] GARCH
• [x] RUGARCH
• [x] RMGARCH

Bayesian Models

Refer to bayesmodels (Website, GitHub)

• [x] Sarima: bayesmodels connects to the bayesforecast package.
• [x] Garch: bayesmodels connects to the bayesforecast package.
• [x] Random Walk (Naive): bayesmodels connects to the bayesforecast package.
• [x] State Space Model: bayesmodels connects to the bayesforecast and bsts packages.
• [x] Stochastic Volatility Model: bayesmodels connects to the bayesforecast package.
• [x] Generalized Additive Models (GAMS): bayesmodels connects to the brms package.
• [x] Adaptive Splines Surface: bayesmodels connects to the BASS package.
• [x] Exponential Smoothing: bayesmodels connects to the Rglt package.

Boosted Models: CatBoost, LightGBM

Refer to boostime (GitHub).

• [x] ARIMA + Catboost
• [x] ARIMA + LightGBM
• [x] Prophet + Catboost
• [x] Prophet + LightGBM

Conformal Predictions

• [x] Investigate integration of probably for Conformal Regression and Confidence Intervals. Follow GH Issue #173

Vector Autoregression (VAR)

• [ ] Investigate integration of VAR models. Follow GH Issue #77.

Hierarchical Time Series

• [ ] Investigate integration of hts package. https://github.com/earowang/hts

Modeltime Targets

• [ ] Develop a reproducible targets workflow for ARIMA, Prophet, and Exponential Smoothing. Incorporate reporting and error diagnostics. #109

Spark: Distributed Compute

Investigate the possibility of adding Spark backend to improve scalability of modeltime, specifically for nested (iterative) forecasting.

• [x] Spark Backend #105 - See The Modeltime Spark Backend

Facebook Neural Prophet

• [ ] Create a Modeltime Neural Prophet Extension.

LinkedIn Greykite

• [ ] Investigate algorithm. Determine if backend has value. https://linkedin.github.io/greykite/docs/0.1.0/html/index.html

Uber Orbit

• [ ] Investigate Uber Orbit library for Bayesian Forecasting. https://github.com/uber/orbit

Facebook Kats

• [ ] investigate Facebook Kats library https://github.com/facebookresearch/Kats

Temporal Hierachical Forecasting (THIEF)

• [x] Implement the thief package for ensembling using temporal hierarchical forecast (THIEF) aggregations. #117

Smooth

Investigate smooth Library. https://github.com/config-i1/smooth The smooth library has several benefits including:

• [x] ETS + XREGS: An es() function that implements Exponential Smoothing with handling of xregs. This can be a new engine in exp_smoothing()
• [x] A Seasonal ARIMA auto.ssarima() / ssarima() implemented in state space. These can be additional options in arima_reg().

These methods have shown promising results in comparison to Facebook Prophet.

Follow #119.

Merlion

• [ ] Investigate Salesforce Merlion for incorporating into Modeltime https://github.com/salesforce/Merlion

Pytorch Forecasting

• [ ] Investigate Pytorch forecasting #163

[edgBR]

Dear @mdancho84,

I see a lot of parallelism between model time and the fable package. Even that fable is designed for multiple time series models I believe that sharing some of the structures could be very beneficial for both projects. I have noticed that for instance you have a model time table. Would it be possible that you look into the mable (mdl_df) class from fable?

BR /E

[mdancho84]

Yes, I agree there are parallels. I'm not sure what specifically to look at compared to fable. I do plan to incorporate most of the forecast package functionality into modeltime. Scalable time series is something else I will look into down the road, but it becomes more complex when tuning is involved.

[edgBR]

Hi @mdancho84

I think the main issue with tunning is that you will have a conflict when you are running fable in parallel as you will not have control over the forked processes that the HPO needs. However I am not an expert in the topic but I for sure understand the challenges here.

However I see now 3 main packages to perform forecasting in R:

• forecastML: for direct forecasting of multiple grouped time series using machine learning and deep learning.

• fable: for recursive forecasting of multiple grouped time series using time series models and prophet

• modeltime: for recursive forecasting using machine learning and deep learning.

forecastML plans to be compatible with fable and there is even a vignette in order to explain reconciliation forecasts.

I think that it would be amazing if all of this packages become part of a modern time series forecasting ecosystem.

BR /E

[mdancho84]

I’m a big fan of fable, I just couldn’t get it to work for anything using machine learning and deep learning.

The tidymodels system is an amazing evolution of caret that is focused on integration and standardization of machine learning. This makes it much easier to control and simplify the vast ML & DL ecosystem. This is why it only takes minimal lines of code to go from idea to execution. Modeltime and timetk simply add a workflow for forecasting on top of the excellent tidymodels ecosystem.

I don't know if there will be one framework that will consolidate all of these great ideas. I agree it would be amazing.

My goal is to first and foremost help those that want to create high-performance forecasts have the tools available to do so in as few lines of code as possible.

[mdancho84]

TBATS Model TBATS is added to the seasonal_reg(), seasonal time series regression model specification.

seasonal_reg(
    seasonal_period_1 = 7,
    seasonal_period_2 = 30,
    seasonal_period_3 = 365
) %>%
    set_engine("tbats")

[mdancho84]

Deep Learning (RNN / LSTM) with Modeltime GluonTS

The modeltime.gluonts library has been created to handle Deep Learning by integrating the Time Series Deep Learning Library, GluonTS. Resources:

• Modeltime GluonTS Docs: https://business-science.github.io/modeltime.gluonts/
• GluonTS Docs: https://ts.gluon.ai/index.html

[spsanderson]

I don't recall if holtwinters works with modeltime, does it?

[mdancho84]

I don't recall if holtwinters works with modeltime, does it?

Yes, Holt-Winters is a special form of ETS. So you can use exp_smoothing() to make a holt winters model.

[mdancho84]

Added Conformal Interval Evaluation #173