The Pendle trading returns are measured on markets where the strategy had already won. Jia, You, Liu, Luo and Sun have produced a dataset worth downloading, but the trading case rests on that selection.
What gets traded?
Pendle splits a yield-bearing token into two instruments. The Principal Token (PT) resembles a zero-coupon bond, redeeming at par at a fixed maturity. The Yield Token (YT) receives the floating yield and decays to zero. The paper describes Pendle as an on-chain interest-rate derivatives protocol and labels PT and YT yield trading derivatives. PT prices determine implied APY; the two move inversely. The proposed trade waits for implied yields to jump, buys discounted PT, then earns from reversion and pull-to-par.
The dataset contains 896,811 hourly rows collected from the Pendle V2 API. Coverage is Ethereum mainnet only, from January 9, 2024 to January 1, 2026: 340 markets, each defined by an underlying and a maturity, across 159 underlyings. Individual markets have 223 to 9,207 rows. Coverage ranges from 89.62% to 100%.
The forecasting target is annualized realized volatility of implied APY over the next 24 hours. It takes the RMS of hourly log changes in APY and scales it by sqrt(8760). Tail-MSE, the authors' loss function, multiplies squared error by alpha whenever a forecast undershoots realized volatility. The tested alpha values are {1, 3, 5, 7, 10}. Training uses six models from the THUML library (Mamba, iTransformer, PatchTST, DLinear, TimesNet, Autoformer), a 96-hour lookback and a chronological 60/20/20 split.
Those forecasts feed the "Tail-MSE Sniper", a PT-only strategy that waits in cash until it forecasts a deep tail event. Comparisons use an MSE-trained twin, buy-and-hold and DCA. The headline is a win count: the strategy earns the highest ROI among four strategies in 181 of 340 markets (53.2%).
We could not verify any of this ourselves. The strategy requires maturity-specific PT prices, implied APY and contract lifecycles. Our data contains none of these instruments. Spot crypto bars offer no substitute because they have neither maturity nor yield decomposition.
Replicators face three judgment calls before getting far. Section 3 reports taxonomy counts of 152/83/37/71, totaling 343; the appendix and backtest table instead report 156/83/36/65, totaling 340. The target uses ln(y_t/y_{t-1}) of implied APY. The authors say y_t > 0 holds for standard yield-bearing assets, yet their field table gives an implied APY range of [-1.00, 60.27]. They also describe funding rates as model inputs, although we did not find them in the data dictionary.
Upward forecasts, undefined tails
Tail-MSE applies an asymmetric squared loss. Increasing alpha pushes the whole forecast upward. The loss specifies neither an extreme event nor a threshold defining one. The abstract's promised "tail risk awareness" therefore amounts to a systematic upward bias across forecasts.
The evaluation metric named Tail-MSE receives conflicting descriptions. Section 5.2 describes weighted error on underestimation; Section 5.3 describes error on tail events. We found neither a formula nor a definition of a tail event. Its alpha = 1 column also separates this metric from the training loss. At alpha = 1, the weighted training loss equals plain MSE, yet Mamba records Tail-MSE of 0.083 and Global MSE of 0.071.
Several claims in the prose conflict with Table 4. The text selects alpha = 3 as optimal for Mamba, citing Tail-MSE of 0.051 versus a 0.083 baseline and a reported 38.8% improvement. But alpha = 7 does better on both metrics: 0.036 tail and 0.038 global. TimesNet's stated 30.9% improvement at 0.046 also runs against the table. Its alpha = 1 result is 0.035, making the weighted model worse. And DLinear contradicts the claim that tail error exceeds global error for every model at alpha = 1: its figures are 0.447 and 0.493, respectively.
Directional accuracy fits the prose least well. The text says every model reaches its peak at alpha = 3 or 5 and reports a 5.2-point average gain. Four of six models actually peak at alpha = 1 in the table: TimesNet (55.42%), PatchTST (53.14%), DLinear (50.00%) and Autoformer (52.91%). Mamba's peak comes at alpha = 7 (54.38%). Only iTransformer reaches its best result in the stated range, recording 52.59% at alpha = 5.
The text also gives 54.7% for iTransformer and TimesNet at alpha = 3. The corresponding table entries are 51.84% and 54.89%. DLinear drops to 27.33% at alpha = 10, well below a coin flip's 50%. The configuration table covers the Mamba runs and specifies a single seed (42), with hyperparameters copied one-for-one from the MSE trials.
The 181 winners out of 340 determine the return sample
The paper claims "superior risk-adjusted returns in over half of all markets." Across the full universe, the supporting result is the count: Tail-MSE achieves the highest ROI among four strategies in 181 of 340 markets. Selection uses ROI. We found neither a Sharpe ratio nor return volatility anywhere in the backtest.
All ROI and drawdown statistics beyond that count depend on the selection. The reported median ROI is 7.8% (IQR 5.2% to 9.6%), versus 2.9% for buy-and-hold. Median drawdown is 3.7%, versus 6.8%. These figures cover the 181 winners, a restriction the authors disclose openly in the appendix. The introduction's contribution list puts the 181-market count beside the 7.8% median ROI and 3.7% drawdown without carrying over that qualification.
The authors defend the subset as a way to isolate market microstructures where tail-risk-aware timing earns "statistically significant alpha". No significance test is reported. Selection happens ex post, and the subset is useful only if the winners can be identified beforehand. The taxonomy win rates provide the paper's only ex-ante handle: 69.2% in stablecoins versus 27.8% in BTC-Fi.
Selecting markets by the winner's ROI guarantees that winner a higher ROI than its alternatives. The excess over buy-and-hold, described as "ranging from 1.77 to 9.50 percentage points", follows from the selection itself. The claim of stochastic dominance in 181 markets asks more of the data than it can support. We did not find ROI or drawdown statistics for the remaining 159 markets.
The sector results weaken the headline further. Tail-MSE wins in 27 of 83 ETH-Beta markets and 10 of 36 BTC-Fi markets. It therefore loses in most markets in the volatile sectors where tail forecasting should help. The paper acknowledges this, defending the result with substantially higher mean ROI when the strategy wins. Those means, 34.16% in ETH-Beta and 25.70% in BTC-Fi, also exclude the losers. Stablecoins deliver its best win rate, at 108 of 156.
The comparisons set a low bar, too. Hold a PT to maturity and it converges to par, leaving buy-and-hold positive in the token's own units. The paper also says DCA often incurs cash drag in upward-trending yield markets.
Turning a volatility forecast into an order
We could not reconstruct this step. The model predicts unsigned volatility of implied APY. A spike could mean falling yields and a PT rally, or rising yields and a PT selloff. The paper says the Sniper "deploys capital only when deep tail events" are forecast, meaning market crashes. We found no threshold, position-sizing rule or explanation of how the strategy infers direction. The MSE twin also changes names, appearing as MSE-Based Swing in Section 6.1 and MSE Sniper in Appendix A.4.
The setup leaves two further gaps. It applies a filter "that isolates markets exhibiting significant volatility", yet reports results for all 340 markets. The backtest runs from January 2024 to January 2026, while model training uses a 60/20/20 split. We did not find a statement identifying which trades belong to the test window.
Execution starts with $10,000, trades daily and assumes zero slippage. We found no fee or gas assumption. Median ETH-Beta TVL declined 98%, from $51.3M in Q1 2024 to $1.1M by late 2025. The zero-slippage assumption thus extends to ETH-Beta pools whose median TVL finished near $1.1M.
The dataset stands on its own.
The trade would merit another look with one table: ROI and drawdown for all four strategies over all 340 markets, restricted to the test window, alongside an entry rule detailed enough to code.