Forecaster

Evaluating and Exploring Timeseries Data with Machine Learning

Time series forecasting has applications across domains and industries, especially in healthcare, but the technical expertise required to analyze data, build models, and interpret results can be a barrier to using these techniques. Forecaster presents a web platform that makes the process of analyzing and plotting data, training forecasting models, and interpreting and viewing results accessible to researchers and clinicians. Users can upload data and generate plots to showcase their variables and the relationships between them. The platform supports multiple forecasting models and training techniques, which are highly customizable according to the user’s needs. Additionally, recommendations and explanations can be generated from a large language model that can help the user choose appropriate parameters for their data and understand the results for each model. The goal is to integrate this platform into learning health systems for continuous data collection and inference from clinical pipelines. Forecaster handles many different aspects of the process of generating predictions in an automated, user-friendly way, making these powerful tools more accessible to clinicians and researchers.

Forecaster is an open-source, web-based platform that incorporates data pre-processing, model training, and results interpretation into a single, easy-to-use interface. After uploading a dataset, users can choose which of their variables to include in the analysis, such as which column should be predicted. Then, users can select what models they want to train and modify the parameters for those models. An LLM assistant is provided that can provide explanations and recommendations on parameter choices and models. It measures dataset size, time periods, and distributions within the dataset.

The Forecaster is built on a modular, dual-backend architecture. The primary frontend web-serving processes, such as user management and database interactions, are performed with a PHP backend as part of a LAPP (Linux, Apache, PostgreSQL, PHP) stack architecture. This system is built using a Model-View-Controller design to appropriately relate processes together. A PostgreSQL relational database is incorporated to store user metadata and overall time series dataset information. CILogon is used for identity management and security of user credentials when accessing the site. A separate Python Flask service was developed and incorporated for technical processes such as analytical processing, job creation, and LLM interactions. This separation allows the platform to leverage robust Python libraries and API interactions for more computationally expensive or complex tasks.

The PHP web service backend interacts with this Python server using the Guzzle HTTP client to make requests to the Flask server’s RESTful API. The LLM is queried using an OpenAI-compatible API. Currently, Forecaster is using the DeepSeek-R1 model to ensure that all data remains secure and private. This model leverages an 8-GPU H200 cluster with 141GB VRAM per GPU to host the models’ weights. Any open-source or commercial reasoning models could be substituted if the user cannot use DeepSeek.

The models the user can select include ARIMA, exponential smoothing, linear regression, XGBoost, random forest, light gradient boosting machine, N-Linear model, and temporal fusion transformer. These models vary in complexity and architecture, so different models will work better for different kinds of datasets.

Four different public datasets were used to test the forecasting performance of the site. These datasets were chosen to ensure a variety of industries, use cases, and dataset types were tested. The datasets for this project were provided by: 

1. Banachewicz K. Norway bicycles [Internet]. 2021. Available from: https://www.kaggle.com/datasets/konradb/norway-bicycles/data
2. Andrikov D. ECG timeseries for prediction [Internet]. 2025. Available from: https://www.kaggle.com/datasets/denisandrikov/ecg-timeseries-for-prediction
3. Anttal TS. Covid 19 dataset till 22/2/2022 [Internet]. 2022. Available from: https://www.kaggle.com/datasets/taranvee/covid-19-dataset-till-2222022
4. Store sales – time series forecasting [Internet]. Available from: https://www.kaggle.com/competitions/store-sales-time-series-forecasting/data?select=train.csv 

Aaron Mullen was the lead developer for Forecaster. A paper detailing further information can be found at https://arxiv.org/abs/2512.07992.

This research was supported in part by the National Institutes of Health under award number UL1TR001998. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

In the current implementation, Forecaster uses the DeepSeek-R1 model. This leverages an 8-GPU H200 cluster with 141GB VRAM per GPU.