Implementation of a Demand Forecasting System Using LSTM, XGBoost, and Temporal Fusion Transformer with a Distributed Data Pipeline

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Yashraj Umesh Panhalkar, Harsha Peshave, Naveed Malik, Prof. Pranali Navghare
Format: Recurso digital
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866901272998707200
author Yashraj Umesh Panhalkar
Harsha Peshave
Naveed Malik
Prof. Pranali Navghare
author_facet Yashraj Umesh Panhalkar
Harsha Peshave
Naveed Malik
Prof. Pranali Navghare
contents This paper presents the end-to-end implementation terms. XGBoost exploits the explicit lag and festival columns of a demand forecasting system for a regional FMCG distributor through regularized tree splits and achieves sub-5 ms inference. in Maharashtra, India. The system comprises a Node.js/Express REST API backed by MongoDB, a Python Flask ML sidecar, TFT routes known future inputs—festival calendar, quarter and a React dashboard. Three machine learning models—Long indicators—through a dedicated encoder, enabling multi-head Short-Term Memory (LSTM), XGBoost, and Temporal Fusion self-attention over prior-year festival quarters. All three share Transformer (TFT)—were trained on real SKU-level quarterly the same feature set, so accuracy differences are attributable sales data. The ML pipeline covers outlier clipping, categorical to architecture alone. encoding, lag feature construction, festival-proximity scoring, and chronological train/validation/test splitting. On the held-out B. Contributions test set, LSTM achieved RMSE = 4,457.49 and MAPE = 18.34%; XGBoost achieved RMSE = 3,526.88 and MAPE = 14.87%; TFT Three aspects distinguish this work. First, the dataset is produced RMSE = 724.44, MAE = 381.67, MAPE = 6.32%, and operational rather than synthetic, carrying real-world noise, WAPE = 4.50%. The backend exposes authenticated REST end- return transactions, and irregular festival timing. Second, fes- points for CSV ingestion, feature preprocessing, model dispatch, and forecast retrieval. This paper documents each implementa- tival proximity is encoded as a continuous score rather than a tion layer—repository structure, data pipeline, model design, API binary flag, capturing the gradual pre-festival demand ramp- routes, and deployment strategy—as a reproducible blueprint for Third, the full system—preprocessing, training, inference, distributor-scale forecasting.
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19878507
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Implementation of a Demand Forecasting System Using LSTM, XGBoost, and Temporal Fusion Transformer with a Distributed Data Pipeline
Yashraj Umesh Panhalkar
Harsha Peshave
Naveed Malik
Prof. Pranali Navghare
Demand forecasting
LSTM
XGBoost
Tempo-ral Fusion Transformer
Node.js
Express
MongoDB
REST API
festival seasonality
FMCG
inventory optimization
This paper presents the end-to-end implementation terms. XGBoost exploits the explicit lag and festival columns of a demand forecasting system for a regional FMCG distributor through regularized tree splits and achieves sub-5 ms inference. in Maharashtra, India. The system comprises a Node.js/Express REST API backed by MongoDB, a Python Flask ML sidecar, TFT routes known future inputs—festival calendar, quarter and a React dashboard. Three machine learning models—Long indicators—through a dedicated encoder, enabling multi-head Short-Term Memory (LSTM), XGBoost, and Temporal Fusion self-attention over prior-year festival quarters. All three share Transformer (TFT)—were trained on real SKU-level quarterly the same feature set, so accuracy differences are attributable sales data. The ML pipeline covers outlier clipping, categorical to architecture alone. encoding, lag feature construction, festival-proximity scoring, and chronological train/validation/test splitting. On the held-out B. Contributions test set, LSTM achieved RMSE = 4,457.49 and MAPE = 18.34%; XGBoost achieved RMSE = 3,526.88 and MAPE = 14.87%; TFT Three aspects distinguish this work. First, the dataset is produced RMSE = 724.44, MAE = 381.67, MAPE = 6.32%, and operational rather than synthetic, carrying real-world noise, WAPE = 4.50%. The backend exposes authenticated REST end- return transactions, and irregular festival timing. Second, fes- points for CSV ingestion, feature preprocessing, model dispatch, and forecast retrieval. This paper documents each implementa- tival proximity is encoded as a continuous score rather than a tion layer—repository structure, data pipeline, model design, API binary flag, capturing the gradual pre-festival demand ramp- routes, and deployment strategy—as a reproducible blueprint for Third, the full system—preprocessing, training, inference, distributor-scale forecasting.
title Implementation of a Demand Forecasting System Using LSTM, XGBoost, and Temporal Fusion Transformer with a Distributed Data Pipeline
topic Demand forecasting
LSTM
XGBoost
Tempo-ral Fusion Transformer
Node.js
Express
MongoDB
REST API
festival seasonality
FMCG
inventory optimization
url https://doi.org/10.5281/zenodo.19878507