Table of Contents: :: Library Catalog

Saved in:

Bibliographic Details
Main Author:	Jha, Abhinav Kumar
Format:	Recurso digital
Language:
Published:	Zenodo 2026
Online Access:	https://doi.org/10.5281/zenodo.19260001
Tags:	Add Tag No Tags, Be the first to tag this record!

Table of Contents:

Customer Lifetime Value (CLV) forecasting is a cor- nerstone of proactive customer relationship manage- ment and resource allocation in modern e-commerce. Traditional approaches, such as probabilistic mod- els (BG/NBD) or classical time-series techniques (ARIMA), struggle with high-cardinality categorical variables, non-linear temporal dynamics, and the in- herent volatility of user behavior. Furthermore, stan- dard deep neural networks typically provide deter- ministic point-estimates, failing to capture the asym- metric business risk associated with overestimating or underestimating future spend. This paper intro- duces a Temporal Fusion Transformer (TFT) archi- tecture, implemented from scratch in TensorFlow, to predict multi-horizon CLV. By integrating Quantile Regression, Variable Selection Networks (VSNs), and Gated Residual Networks (GRNs), the model out- puts an 80% prediction interval (P10, P50, P90), offer- ing calibrated uncertainty estimates for future spend- ing. To validate practical utility and reproducibility, the architecture is deployed as an open-source, end- to-end full-stack scalable inference pipeline. Experi- mental results on synthetic e-commerce data demon- strate that the TFT significantly outperforms stan- dard Long Short-Term Memory (LSTM) baselines in both aggregate pinball loss and temporal inter- pretability.

Similar Items