Supriyadi
Sekolah Tinggi Ilmu Ekonomi Enam Enam Kendari

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

Zero-Inflated Asymmetric Rate-Distortion Quantization: Risk-Aware Mixed-Precision Online Inference for Real-Time Tax Revenue Forecasting Nita Hasnita; Muhammad Faried Pratama; Supriyadi
International Journal of Management, Economic and Accounting Vol. 3 No. 3 (2025): June 2025
Publisher : Yayasan Multidimensi Kreatif

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.61306/ijmea.v3i3.838

Abstract

Timely tax revenue forecasts support fiscal planning, yet deploying probabilistic forecasting models in operational systems requires a difficult balance between predictive fidelity and computational speed. That balance is particularly consequential when tax collections contain frequent zeros, extreme values, and economically unequal consequences of forecast errors. This study develops a zero-inflated asymmetric rate-distortion quantization framework for mixed-precision inference with a pretrained quantile Temporal Fusion Transformer. Its objective combines an asymmetric Huber-based quantile penalty, a zero-revenue penalty, and additional weight on distributional tails. A primal-dual controller adjusts numerical precision across model units in response to recent filing patterns and a profiled latency constraint. The source manuscript reports normalized distortion values of 0.734 and 0.762 for the proposed method on its value-added-tax and FRED-derived evaluation streams, respectively, compared with a uniform-quantization reference of 1.000; its reported latency-feasibility rate is 0.981. These values are reproduced here as reported, not independently validated; the source does not supply executable code, complete provenance of the underlying data, or hardware benchmark logs. The proposed approach offers a testable route toward aligning inference resources with fiscal forecasting risk, but replication and deployment-level validation remain necessary.