Skip to content

Towards Event-Aware Forecasting in DeFi: Insights from On-chain Automated Market Maker Protocols

Home / Papers / Towards Event-Aware Forecasting in DeFi: Insights from On-chain Automated Market Maker Protocols

Towards Event-Aware Forecasting in DeFi: Insights from On-chain Automated Market Maker Protocols

Authors: Huaiyu Jia, Jiehshun You, Yizhi Luo, Jingyu Liu, Shuo Sun Date: 2026-04-22 Paper ID: openalex:2604.20374

Summary

This paper addresses the unique event-driven nature of price discovery in Automated Market Maker (AMM) protocols by analyzing on-chain swap and reserve ratio events. The authors curate a large-scale, fine-grained dataset of 8.9 million events from major protocols and propose an Uncertainty Weighted Mean Squared Error (UWM) loss function for Time-Point Process (TPP) modeling. The UWM loss effectively integrates block interval regression into traditional TPP objectives, significantly improving temporal prediction accuracy while maintaining event classification robustness in DeFi environments.

Key Contributions

  • Constructed a fine-grained dataset of 8.9 million on-chain event records across four major AMM protocols (Pendle, Uniswap v3, Aave, Morpho).
  • Proposed the Uncertainty Weighted Mean Squared Error (UWM) loss function to improve time-point process modeling of event-driven DeFi dynamics.
  • Demonstrated that UWM reduces time prediction error by 56.41% compared to standard TPP objectives on eight baseline architectures.

Open Questions & Future Work

Key Concepts

Archivist Review

I approved the Uncertainty Weighted Mean Squared Error loss as a reusable mechanism for TPP modeling, particularly for event-driven time series. I also approved the identified research direction as a bottleneck, ensuring it fits established naming conventions. I rejected the individual protocol datasets as they are components of a single aggregated collection rather than foundational, standalone research benchmarks.

Approved Concepts

  • Uncertainty Weighted Mean Squared Error: This loss function addresses the challenge of predicting discrete on-chain events by balancing event types with precise temporal (block interval) regression.

Approved Open Questions

  • Event-Aware DeFi Forecasting Bottleneck: This is fundamental to moving beyond classical finance paradigms and creating accurate, high-frequency forecasting tools tailored to the unique mechanics of blockchain-native liquidity pools.

Rejected Candidates

  • [open_question] Event-Aware Forecasting in DeFi (event-aware-defi-forecasting-frameworks) - other: Renamed to match vault convention as a bottleneck.

Metadata & Links

url
https://arxiv.org/abs/2604.20374
paper_id
2604.20374
paper_source
openalex
domain
time-series
tags
time-seriesforecasting
architectures
datasets
concept_slugs
uncertainty-weighted-mean-squared-error
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-25T05:38:02Z