A Multi-Asset Dataset and Tail-Risk Framework for DeFi Yield Trading Derivatives (arXiv (Cornell University))

Summary

• The authors construct a benchmark dataset spanning 24 months from January 2024 to January 2026 containing 896,811 hourly observations across 340 active markets.p.2
• Analysis of capital allocation shows BTC-Fi markets reached a median TVL of $16.07M with a peak of $93.3M in Q3 2024, while the ETH-Beta sector suffered a 98% contraction in median liquidity.p.5
• Experimental evaluation demonstrates that standard loss functions underperform on tail events, whereas the proposed Tail-MSE achieves up to a 40.0% improvement in capturing extreme volatility for Autoformer models.p.7
• The Tail-MSE Sniper strategy established superior risk-adjusted returns by winning in 181 out of 340 markets, achieving a 53.2% overall win rate.p.8
• Backtesting across winning markets shows the Tail-MSE strategy achieved a median ROI of 7.8% and a median maximum drawdown of 3.7%, compared to a 2.9% median ROI and 6.8% maximum drawdown for Buy-and-Hold.p.15

Key insights

1. The research introduces the first fine-grained, large-scale dataset for Pendle yield trading derivatives spanning 340 markets over two years.p.1
2. Standard MSE loss functions treat error directions symmetrically, exposing yield trading models to catastrophic downside risk during extreme market shifts.p.2
3. Normalized Principal Token prices reveal a tiered discount hierarchy where low-volatility BTC-Fi assets stay close to par with a median ratio of 0.9937.p.4
4. Return distributions empirically confirm a pull-to-par effect across all asset categories as Principal Tokens deterministically appreciate to face value at maturity.p.5
5. Incorporating the Tail-MSE loss function raises model directional prediction accuracy to between 52% and 55% across tested architectures.p.8
6. Algorithmic backtesting shows the Tail-MSE strategy achieves stochastic dominance in 53.2% of all analyzed Pendle yield markets.p.8
7. The Tail-MSE approach effectively limits downside risk, maintaining a median maximum drawdown of 3.7% compared to 6.8% for a passive Buy-and-Hold strategy.p.15