Saved in:
| Main Author: | |
|---|---|
| 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!
|
| _version_ | 1866901236788232192 |
|---|---|
| author | Jha, Abhinav Kumar |
| author_facet | Jha, Abhinav Kumar |
| contents | <p><br>Customer Lifetime Value (CLV) forecasting is a cor-<br>nerstone of proactive customer relationship manage-<br>ment and resource allocation in modern e-commerce.<br>Traditional approaches, such as probabilistic mod-<br>els (BG/NBD) or classical time-series techniques<br>(ARIMA), struggle with high-cardinality categorical<br>variables, non-linear temporal dynamics, and the in-<br>herent volatility of user behavior. Furthermore, stan-<br>dard deep neural networks typically provide deter-<br>ministic point-estimates, failing to capture the asym-<br>metric business risk associated with overestimating<br>or underestimating future spend. This paper intro-<br>duces a Temporal Fusion Transformer (TFT) archi-<br>tecture, implemented from scratch in TensorFlow, to<br>predict multi-horizon CLV. By integrating Quantile<br>Regression, Variable Selection Networks (VSNs), and<br>Gated Residual Networks (GRNs), the model out-<br>puts an 80% prediction interval (P10, P50, P90), offer-<br>ing calibrated uncertainty estimates for future spend-<br>ing. To validate practical utility and reproducibility,<br>the architecture is deployed as an open-source, end-<br>to-end full-stack scalable inference pipeline. Experi-<br>mental results on synthetic e-commerce data demon-<br>strate that the TFT significantly outperforms stan-<br>dard Long Short-Term Memory (LSTM) baselines<br>in both aggregate pinball loss and temporal inter-<br>pretability.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19260001 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Temporal Fusion Transformers for Multi-Horizon Probabilistic Customer Lifetime Value Forecasting: An End-to-End Approach Jha, Abhinav Kumar <p><br>Customer Lifetime Value (CLV) forecasting is a cor-<br>nerstone of proactive customer relationship manage-<br>ment and resource allocation in modern e-commerce.<br>Traditional approaches, such as probabilistic mod-<br>els (BG/NBD) or classical time-series techniques<br>(ARIMA), struggle with high-cardinality categorical<br>variables, non-linear temporal dynamics, and the in-<br>herent volatility of user behavior. Furthermore, stan-<br>dard deep neural networks typically provide deter-<br>ministic point-estimates, failing to capture the asym-<br>metric business risk associated with overestimating<br>or underestimating future spend. This paper intro-<br>duces a Temporal Fusion Transformer (TFT) archi-<br>tecture, implemented from scratch in TensorFlow, to<br>predict multi-horizon CLV. By integrating Quantile<br>Regression, Variable Selection Networks (VSNs), and<br>Gated Residual Networks (GRNs), the model out-<br>puts an 80% prediction interval (P10, P50, P90), offer-<br>ing calibrated uncertainty estimates for future spend-<br>ing. To validate practical utility and reproducibility,<br>the architecture is deployed as an open-source, end-<br>to-end full-stack scalable inference pipeline. Experi-<br>mental results on synthetic e-commerce data demon-<br>strate that the TFT significantly outperforms stan-<br>dard Long Short-Term Memory (LSTM) baselines<br>in both aggregate pinball loss and temporal inter-<br>pretability.</p> |
| title | Temporal Fusion Transformers for Multi-Horizon Probabilistic Customer Lifetime Value Forecasting: An End-to-End Approach |
| url | https://doi.org/10.5281/zenodo.19260001 |