A Multi-Asset Dataset and Tail-Risk Framework for DeFi Yield Trading Derivatives

• Yield-bearing assets split into principal and yield streams, enabling passive income or active yield speculation while enhancing capital efficiency.The introduction motivates forecasting because PT price and implied APY move in opposite directions.Source: §1 · p. 1
• Prior datasets largely focus on spot markets and omit yield trading’s complete lifecycle, constraining empirical analysis of market behavior.The missing observations include yield decomposition mechanics and token time-decay dynamics.Source: §1 · p. 1
• Pendle’s fine-grained dataset covers 24 months and 340 markets with hourly snapshots, supporting a comprehensive empirical study of yield trading.Pendle is selected because it has sustained activity and substantial trading volume, unlike other protocols described as unsuitable for systematic study.Source: §1 · p. 1
• Tail-MSE uses asymmetric dynamic penalties to emphasize extreme downside risks and volatility spikes in heavy-tailed yield data.The authors evaluate the loss across six time-series models.Source: §1 · p. 2
• The Tail-MSE Sniper outperforms baselines in 181 markets (53.2%), with 7.8% median ROI and 3.7% maximum drawdown.It also achieves a 69.2% win rate in Stablecoin markets.Source: §1 · p. 2

2.1 DeFi Yield Derivatives

• DeFi research has shifted from asset-level price analysis toward protocol-level yield dynamics, including empirical work on Uniswap V3 liquidity and Aave lending.Earlier work also examined cryptocurrencies, perpetual futures, Meme tokens, and on-chain transaction data.Source: §2.1 · p. 2
• Research on yield derivatives remains sparse and largely theoretical, lacking fine-grained data-driven empirical analysis; Pendle is notable for sustained market activity.The passage notes that many emerging yield protocols have collapsed amid flawed economic models or liquidity crises.Source: §2.1 · p. 2

2.2 Quantitative Strategies in DeFi

• Quantitative research rarely models or trades on-chain protocol yield, leaving Pendle’s dual-token pricing dynamics and yield-trading strategy opportunities underexplored.The paper positions its customized strategy as addressing this gap with fine-grained protocol data.Source: §2.2 · p. 2

3.1 Dataset Formulation; 3.2.1 Data Collection and Scope.

• Pendle decomposes standardized underlying assets into PT and YT, with PT converging to one and YT decaying to zero at maturity.The dataset covers Ethereum mainnet from January 2024 to January 2026, with 340 products and hourly protocol-level and PT/YT asset-level data aligned by contract address and time.Source: §3.1 · p. 3, §3.2.1 · p. 3

3.3.1 Asset Class Distribution.

• The 340 markets fall into four classes: Stablecoins (152, 44.3%), ETH-Beta (83, 24.2%), BTC-Fi (37, 10.8%), and Others (71, 20.7%).The taxonomy groups markets by underlying yield source, including rate-stable collateral, ETH-linked restaking, and BTC-linked positions.Source: §3.3.1 · p. 3

3.3.2 Temporal Evolution of Pendle Markets Creation.

• ETH-Beta market creation surged during the 2024 restaking cycle, while 2025 brought BTC-Fi rotation and a stablecoin rebound described as a flight to quality.Figure 5 tracks monthly creation volume by asset class; the two phases are tied respectively to airdrop incentives and their end.Source: §3.3.2 · pp. 3–4, Figure 5 · p. 3

3.3.3 Product Lifecycle Analysis.

• ETH-Beta markets have the longest median tenor at around 114 days, with some exceeding 350 days; BTC-Fi and stablecoin lifetimes cluster around 100 days.The section links ETH-Beta tenor to restaking lockups and shorter BTC-Fi and stablecoin lifetimes to liquidity incentive cycles.Source: §3.3.3 · p. 4

3.4.2 PT Normalized Price Structures.

• BTC-Fi PTs trade near par (median 0.9937), while ETH-Beta and Others show larger discounts (median ∼0.94–0.95); stablecoin PT prices reached 0.955 in Q1/Q3 2025.The section interprets the stablecoin price elevation as signaling rotation toward stable-return products.Source: §3.4.2 · p. 4

3.4.3 Token Supply Strategies.

• Stablecoins have the highest and most consistent circulating supply (median ∼10^6), while BTC-Fi and ETH-Beta medians are ∼10^1 and 10^2, respectively.The analysis also reports declining median ETH supply over time and a heterogeneous Others range of 10^3–10^9.Source: §3.4.3 · pp. 4–5

3.4.4 Market Capitalization and Capital Allocation.

• BTC-Fi markets had the highest median TVL ($16.07M), while ETH-Beta median liquidity contracted 98%, from $51.3M in Q1 2024 to $1.1M in late 2025.Stablecoin TVL remained consistent at a median of $7.12M; BTC-Fi peaked at $93.3M in Q3 2024.Source: §3.4.4 · p. 5

3.4.5 Return Distributions.

• Stablecoin 30-day PT price changes show minimal variance, whereas BTC-Fi has the widest dispersion and ETH-Beta follows a smoother, higher-volatility curve.All categories show a positive median shift, which the analysis attributes to PT appreciation toward face value.Source: §3.4.5 · p. 5, Figure 7 · p. 5

4.1 Problem Description

• The prediction task forecasts annualized realized volatility over a future H-hour horizon from historical implied-APY log returns.Annualization accounts for DeFi markets operating 8,760 hours per year.Source: §4.1 · p. 5
• The method treats predicted realized volatility as a fair-volatility benchmark, with deviations from market-implied volatility identifying potential volatility-arbitrage opportunities.The paper frames this as a source of alpha compared with directional price prediction in efficient markets.Source: §4.1 · p. 5

4.2 Risk-Aware Asymmetric Loss Function

• Tail-MSE weights squared errors asymmetrically, heavily penalizing volatility underestimation because it risks ruin while overestimation incurs opportunity cost.The dynamic weight is α for underestimation and 1 for overestimation, with α representing the risk-aversion coefficient.Source: §4.2 · pp. 5–6

4.3 Temporal Neural Network Architecture

• A unified architecture processes yield-series log returns into normalized windows and embeddings, then uses stacked temporal encoding layers to capture multi-scale volatility dynamics.The framework accommodates multiple time-series models through consistent input-output interfaces.Source: Figure 8 · p. 6, §4.3 · p. 6
• The final hidden state is linearly projected into raw volatility predictions, from which the last H steps are selected, denormalized, and annualized.Source: §4.3 · p. 6

5.1 Setup

• The study forecasts annualized realized volatility from pool-level liquidity, price, TVL, funding-rate, and market-identifier inputs, using chronological 60/20/20 splits to prevent look-ahead bias.Hourly sliding windows are used, and feature standardization uses training-split statistics only.Source: §5.1 · p. 6

5.2 Baselines and Evaluation Metrics

• Six state-of-the-art forecasting architectures are evaluated under standard MSE and asymmetric Tail-MSE, with direction accuracy measuring APY increase/decrease prediction.Tail-MSE dynamically weights underestimation errors using risk-aversion coefficients α∈{1, 3, 5, 7, 10}.Source: §5.2 · pp. 6–7

5.3 Results

• Under standard training (α=1), every model has higher tail-event than global prediction error; increasing α changes that relationship.The experiments characterize this as evidence of asymmetric tail characteristics in Pendle yield data.Source: §5.3 · p. 7
• Mamba at α=3.0 records Global MSE=0.056 and Tail-MSE=0.051, a 38.8% Tail-MSE improvement over baseline; iTransformer records 0.041 and 0.037, a 14.9% improvement.The iTransformer result maintains global prediction accuracy.Source: §5.3 · p. 7
• PatchTST at α=5.0 reaches Tail-MSE 0.083, a 5.8% improvement; DLinear at α=5.0 reaches 0.368, a 17.7% improvement despite higher Global MSE.TimesNet at α=3.0 reaches Tail-MSE 0.046, a 30.9% improvement despite Global MSE rising to 0.054.Source: §5.3 · p. 7
• Autoformer at α=5.0 reaches Tail-MSE 0.134, a 40.0% improvement over baseline.Across architectures, the optimal α varies but falls within 1.0–5.0; α=3.0 and α=5.0 offer the best trade-off between tail-risk mitigation and global accuracy.Source: §5.3 · pp. 7–8
• All models reach 52–55% directional accuracy at α=3.0 and α=5.0, averaging 5.2 percentage points above baseline.iTransformer and TimesNet each reach 54.7% at α=3.0.Source: Figure 10 · p. 7, §5.3 · p. 8

6.1 Practical Applications

• The Tail-MSE Sniper conserves liquidity during market noise and deploys capital when deep tail events are forecast.The strategy is evaluated on PTs, whose deterministic convergence to par at maturity provides a controlled setting for entry-timing tests during market turbulence.Source: §6.1 · p. 8, §6 · p. 8

6.2 Experiments and Strategic Implications

• The backtest starts with $10,000, rebalances daily using Mamba-based MSE or Tail-MSE signals, and filters for markets with significant volatility.Results are aggregated across all 340 markets by token-name taxonomy and compared on ROI and drawdowns.Source: §6.2 · p. 8
• Tail-MSE wins in 181 of 340 markets (53.2%) and exceeds Buy-and-Hold by 1.77–9.50 percentage points in average excess ROI across every taxonomy.Stablecoins have a 69.2% win rate, versus 32.5% for ETH-Beta and 27.8% for BTC-Fi; Others reaches 55.4%.Source: §6.2 · p. 8, Table 1 · p. 8, Figure 11 · p. 8
• The authors report that predictive strategies outperform naive baselines and that Tail-MSE captures black-swan yield spikes.They identify Tail-MSE Sniper as suitable for active traders seeking alpha in volatile rate environments.Source: §6.2 · p. 8
• For stablecoins, the authors recommend Tail-MSE Sniper to target temporary yield spikes and minor de-pegs; for volatile assets, they recommend predictive-model-enhanced DCA.For ETH-Beta, BTC-Fi, and Others, the DCA approach weights purchases toward dips rather than pursuing precise timing.Source: §6.2 · p. 8

7 CONCLUSION

• Across 24 months and 340 markets, the dataset reveals a transition from speculative ETH-Beta dominance to mature, risk-neutral Stablecoin preference; Tail-MSE reduces Mamba error 38.8% and improves Autoformer tail-event capture 40.0% versus baselines.The Tail-MSE-driven strategy achieved stochastic dominance in 53.2% of all markets and a 69.2% win rate in the Stablecoin sector, with consistent excess returns.Source: §7 · p. 9

A.1 Pendle

• Pendle enables tokenization and separate trading of future yield streams.The section introduces a protocol deconstruction focused on its financial engineering and yield-discovery mechanisms.Source: §A.1 · p. 11

A.1.1 Standardization and Structural Decomposition (SY Architecture).

• SY wraps heterogeneous yield-bearing assets, including rebasing stETH and LP tokens, in a unified accounting interface.This standardization supports composability across different yield mechanics.Source: §A.1.1 · p. 11
• The SY layer enables yield mechanics to be algorithmically divided into Principal Tokens (PT) and Yield Tokens (YT).PT represents the asset corpus, while YT represents the right to its income.Source: §A.1.1 · p. 11
• The paper identifies composability across diverse yield mechanics as a key role of SY standardization.Source: §A.1.1 · p. 11

A.1.2 Derivative Pricing Dynamics.

• The no-arbitrage condition sets the SY price equal to the sum of PT and YT prices.Source: §A.1.2 · p. 11
• PT represents fixed-yield exposure and must converge to 1.0 underlying asset at maturity, with its return accruing through price appreciation.Before maturity, its price moves inversely with market interest rates; buying at 0.9 yields approximately 11% upon redemption at 1.0.Source: §A.1.2 · p. 11
• YT represents uncertain future yield cash flows and decays toward zero as maturity approaches and the collection window closes.A YT holder profits only if realized yield exceeds the implied yield paid upfront.Source: §A.1.2 · p. 11

A.1.3 Strategic Utility Landscape.

• PT buyers seek deterministic fixed-yield exposure, YT buyers seek leverage to expected yield surges, and liquidity providers earn fees and incentives while bearing impermanent-loss risk.Source: §A.1.3 · p. 11

A.2 Detailed Dataset

• The dataset covers 340 Pendle liquidity pools over 24 months, with 896,811 hourly observations, and is presented as a first large-scale, multi-asset structured yield-derivative collection.Its records include hourly yield, liquidity, valuation, volatility, and activity measures, with realized implied-yield volatility as the primary prediction target.Source: §A.2 · p. 11, §A.2.2 · p. 12

A.2.1 Taxonomy and Asset Distribution.

• The four asset taxonomies distinguish stablecoin yield profiles, ETH-linked staking and restaking yields, emerging BTC-Fi yields, and heterogeneous long-tail assets.Stablecoins form the risk-neutral baseline with mean-reverting underlying prices; ETH-Beta has complex yield stacking, BTC-Fi carries novelty premiums, and Others spans widely varying yield drivers.Source: §A.2.1 · pp. 11–12

A.2.3 Towards Invest Evaluation.

• BTC-Fi is described as highest in profitability and diversity, ETH-Beta as most universal and reliable, stablecoins as defensive with minimal extreme tail risk, and Others as diverse but erratic.The paper links these profiles respectively to forecasting needs, benchmark-relative farming, principal-preserving cash flow, and tactical short-term volatility arbitrage.Source: Figure 12 · p. 12, §A.2.3 · pp. 12–13
• The proposed Risk-Aware Layered Allocation assigns 50–60% to ETH-Beta and stablecoins, 20–30% to BTC-Fi under specified volatility conditions, and 10% to Others for tactical arbitrage.Source: §A.2.3 · p. 13

A.4 Insights from Trading

• Tail-MSE achieves the highest ROI among the four compared strategies in 181 of 340 markets (53.2%), with a 7.8% median ROI and 3.7% median maximum drawdown, versus 2.9% and 6.8% for B&H.DCA has the lowest median maximum drawdown (2.4%) but a 1.8% median ROI.Source: Figure 13 · p. 15, §A.4 · p. 15
• Stablecoins and Others have the highest Tail-MSE win rates (69.2% and 55.4%), while ETH-Beta and BTC-Fi win less often (32.5% and 27.8%) but yield higher conditional ROI when winning.The passage characterizes opportunities in crypto-native assets as scarcer but associated with wider discount dispersion.Source: §A.4 · p. 15
• In ETH-Beta and BTC-Fi, Tail-MSE captures transient panic-selling discounts; in stablecoins, its narrower but consistent edge comes from avoiding premature entries during minor de-pegging events.The paper recommends steady carry enhancement for stablecoin assets and opportunistic allocations to BTC/ETH derivatives; Table 5 reports excess returns of 1.69%–3.27% for syrupUSDC and USDe, and 10.59% and 6.84% for rsETH and eBTC, respectively.Source: Figure 14 · p. 15, §A.4 · p. 15, Table 5 · p. 15