microsoft/finnts

Microsoft Finance Time Series Forecasting Framework (FinnTS) is a forecasting package that utilizes cutting-edge time series forecasting and parallelization on the cloud to produce accurate forecasts for financial data.

Microsoft Finance Time Series Forecasting Framework

The Microsoft Finance Time Series Forecasting Framework, aka finnts or Finn, is an automated forecasting framework for producing financial forecasts. While it was built for corporate finance activities, it can easily expand to any time series forecasting problem!

• Automated feature engineering, feature selection, back testing, and model selection.
• Access to 25+ models. Both univariate and multivariate models.
• Azure integration to run thousands of time series in parallel within the cloud.
• Supports daily, weekly, monthly, quarterly, and yearly forecasts.
• Handles external regressors, either purely historical or historical+future values.

Installation

CRAN version

install.packages("finnts")

Development version

To get a bug fix or to use a feature from the development version, you can install the development version of finnts from GitHub.

# install.packages("devtools")
devtools::install_github("microsoft/finnts")

Usage

library(finnts)

# prepare historical data
hist_data <- timetk::m4_monthly %>%
  dplyr::rename(Date = date) %>%
  dplyr::mutate(id = as.character(id))

# call main finnts modeling function
finn_output <- forecast_time_series(
  input_data = hist_data,
  combo_variables = c("id"),
  target_variable = "value",
  date_type = "month",
  forecast_horizon = 3,
  back_test_scenarios = 6, 
  models_to_run = c("arima", "ets"), 
  run_global_models = FALSE, 
  run_model_parallel = FALSE
)

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

finnts's Issues

add lightgbm model as default model in Finn

namespace export issue

forecast_time_series function not correctly exported to namespace file.

if ensemble models turned off, only run as many back tests as back test scenarios

bug: ensemble training data creation for yearly forecasts

need to adjust back testing process when running a yearly forecasts. Currently a year is the minimum smallest amount of training data when doing model refitting, which works for quarterly down to daily forecasts. But for yearly it should be something like 3-5 years. Or have an automated process that turns ensembles off if the user doesn't have more than "x" amount of historical years of data.

add github info on documentation site

make it easy for someone who comes to microsoft.github.io/finnts site to view the source code on github.

add NEWS.md to show version history

allow xregs in hierarchical forecast

Refactor code to split modules

Refactoring to finalize the code for first launch

more control over num_cores ran in parallel

By default, only use 2/3 of available CPU cores during any parallel processing. This will help limit the RAM memory errors that might arise. Also enable a user to enter their own number of cores to use, "num_cores" parameter in "forecast_time_series".

document main functions in package

Add CI/CD github actions into package

categorical variable handling

See what other alternatives to dummy columns can be used to reduce data complexity. Should they just be mapped to certain levels for tree based models? How would that effect more simple linear regression models? What about embeddings?

construct_forecast_models bug

the below code results in an error.

library(finnts)

hist_data <- timetk::m4_monthly %>%
  dplyr::filter(date >= "2010-01-01") %>%
  dplyr::rename(Date = date) %>%
  dplyr::mutate(id = as.character(id))

finn_output <- forecast_time_series(
  input_data = hist_data,
  combo_variables = c("id"),
  target_variable = "value",
  date_type = "month",
  forecast_horizon = 12,
  models_to_run = c("arima", "ets"))

Error in construct_forecast_models(full_data_tbl, external_regressors, :
object 'init_azure_batch_parallel_within' not found

don't think the run_model_parallel feature is working properly. Currently we only pass the "init_azure_batch_parallel_within" and "exit_azure_batch_parallel_within" functions as inputs to the contruct_forecast_models function as inputs for "parallel_init_function" and "parallel_exit_func". I don't see a function that just creates a parallel environment to allow run_model_parallel to work properly when parallel_processing is set to "none".

I think a quick fix would be to add to the init_azure_parallel_within and exit_azure_parallel_within functions within azure_batch_parallel.R file to be flexible to having run_model_parallel work in azure and on a local machine when parallel_processing is set to "none". That new function could live in the general_parallel.R file. The new function names could be something like init_parallel_within and exit_parallel_within, and only contain an argument related to parallel_processing choice given by the user.

What do you think @AKannanMSFT?

add bayesian models from bayesmodels package

https://albertoalmuinha.github.io/bayesmodels/

• Sarima
• Garch
• Random Walk (Naive)
• Sate Space Model
• Stochastic Volatility Model
• GAMS
• Adaptive Splines Surface
• Exponential Smoothing

historical data issue on future forecast output

future forecast output can contain both historical values and future forecast values.

If input data contains dates after "hist_end_date" input, remove them when producing the future forecast output. That way the user only sees historical values up until his_end_date, and then sees a future forecast for future periods after input value. Right now we are seeing both forecast and hist values.

research use case and impact of featureterminator package for removing non-important features

https://insidebigdata.com/2021/08/01/video-highlights-featureterminator-package-for-r/

azure machine learning service integration

allow for logging within Azure ML Service.

Log run results

• Finn arguments
• accuracy by combo
• best model chosen by combo

Should allow for better MLOps processes and tracking.

check that Finn 2.0 can run on docker container with necessary software

add more granular update messages as Finn runs

ensure that grouped hierarchical forecast works correctly

turn repo into a github pages site to host documentation and resources to learn more about Finn

add more tests to check various use cases of Finn

incorporate dimensionality reduction techniques

• PCA
• Any others

improve hyperparameter process

Ensure that the optimal hyperparameters are chosen in a way that ensures the most efficient run time and accuracy.

Incorporate new feedback to ensure no data leakage when selecting hyperparameters. Also remove duplication when selecting hyperparameters and refitting models.

Allow user to define how many iterations of hyperparameters to try in tune_grid

prophet-xregs and nnetar-xregs don't work in azure batch

error when running grouped_hierarchy for forecast_approach

Has to do when reconciling to lowest level at the end.

Bug in R1 feature engineering process

Target lags are not working correctly. Lags are created for periods less than forecast horizon.

add racing methods to hyperparameter tuning process

leverage fintune package from tidymodels. Helps limit the amount of hyperparameter combinations to test, speeding up model run time.

https://finetune.tidymodels.org/reference/tune_race_anova.html

https://juliasilge.com/blog/baseball-racing/

final release tracker code scan

Cross-Validation and Back Testing

rethink how hyperparameters and back test folds are ran. Should we go back to Finn 1.0 approach? Or only have a single CV/back test process?

initial package startup messages

• Where dev package lives.
• How to install necessary python resources
• potential conflicts and if package dependencies are out of date

add Neural Prophet into core model list

https://github.com/AlbertoAlmuinha/neuralprophet

add new deep learning models from modeltime.gluonts

https://business-science.github.io/modeltime.gluonts/news/index.html

• pytorch deep_ar
• gp forecaster
• deep state
• update install_gluonts to include torch backend

add embeddings for all data models

capture relationships between categorical data like time series ID and other groupings. Helpful in deep learning models, not sure if helpful in standard multivariate ML models.

Create embeddings from deep learning models. Then use those values for categorical variables instead of using one-hot encoding/dummy variables for categorical data. Woohoo!

Could also create a separate recipe to do this or do some initial testing on the dataset and see if we should switch over to it as default. Maybe make a global option to either use dummy variables or embeddings for categorical data.

Excerpt from fast.ai deep learning for coders book.

Already looks like an easy integration into a recipe. https://embed.tidymodels.org/reference/step_embed.html

Lastly, we need to determine that size of our embedding. There is no steadfast rule on how to do this but a good heuristic given by Jermey Howard of Fast.Ai is to take half the number of unique values then add one.

Documentation Vignettes

• Models used in Finn
• Feature Engineering
• Back Testing and Hyperparameter Tuning
• External Regressors
• Best Model Selection
• Hierarchical Forecasting
• Parallel Processing and Azure Batch
• Production with Azure Machine Learning Pipelines
• Quick Start Guide

Automate the GitHub Pages using GitHub Actions

We need a way to automatically do CICD for the GitHub Pages

break up monolith

Break out into as many smaller functions as possible. That way finnts becomes more granular and less monolithic. Also see how we can make each sub function exported so it can be used by users outside of "forecast_time_series" function.

feasibility of using dtplyr with data.tables for data transformations

Investigate the upside of leveraging data.table transformations to speed up data processing within the package. Could be helpful when forecasting large datasets (millions of rows).

https://dtplyr.tidyverse.org/
https://www.business-science.io/code-tools/2021/04/13/dtplyr-datatable-dplyr-backend.html?utm_content=buffere2297&utm_medium=social&utm_source=linkedin.com&utm_campaign=buffer

local_machine parallel processing bug

objects are not exporting correctly to parallel cluster on local machine.

more control over running global and local models

Enable a user to turn off all global models that run over the entire dataset (if input_data has more than 1 time series) and also be able to turn off individual models that run over each induvial time series (if input_data has more than 1 time series)

create finnts package logo

allow for long form text data to be used as features

automated text feature engineering with NLP

Update Readme.md and other assets to be consistent with branding

Change all labels within the Readme.md and other places to reflect the "Microsoft Finance Time Series Forecasting Framework (FinnTS) " branding.

update description file

• add all contributors (chief economist team and FD&E intern) to authors section
• update emails
• add package description
• correct version number
• change any version issues with dependencies

move repo to public and share with external customers

3M
Nike
Astra Zeneca
Pepsico

Investigate Furrr package for parallel data cleaning

https://www.business-science.io/r/2021/09/14/parallel-processing-furrr.html?utm_content=buffer49022&utm_medium=social&utm_source=linkedin.com&utm_campaign=buffer

Python Support

should we have separate R and Python components to repo if we end up developing a python version of Finn?

@AKannanMSFT something to think about...

dealing with missing data

When filling in missing data (inputed value or with zero), make another binary column that calls out that specific period where data was missing (for target variable and xregs).

add EDA functions to help in data exploration before running finn

Leverage packages like "DataExplorer" to created automated EDA HTML reports. Also make it easy to make decisions around when to choose to leverage things like PCA or remove features that are correlated.

log transform feature

transform back to original data before selecting best model. Also think if that should happen even before model averaging is done.