Pith. sign in

REVIEW 3 major objections 4 minor

Benchmarking Classical and Quantum Models for DeFi Yield Prediction on Curve Finance

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Tree ensembles are the best of six tested models for predicting Curve Finance pool yields, while both quantum models score below 50 percent directional accuracy.

desk verdict A modest but honest benchmark on Curve Finance yield prediction; the headline result is plausible, but the abstract alone cannot verify a fair quantum vs. classical comparison, and the sub-50% quantum accuracy smells like an implementation issue. read the letter →

arxiv 2508.02685 v1 pith:SG6M4M6V submitted 2025-07-22 q-fin.ST cs.LGq-fin.TR

classification q-fin.STcs.LGq-fin.TR
keywords DecentralizedfinanceYieldpredictionCurveXGBoostRandomForestQuantummachinelearningBenchmarkingTimeseriesforecasting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that, on one year of historical data from 28 Curve Finance liquidity pools, classical tree ensembles are the most accurate predictors of DeFi yields among six tested models. XGBoost reaches 71.57 percent directional accuracy with a test MAE of 1.80, Random Forest reaches the lowest test MAE of 1.77 with 71.36 percent accuracy, and both quantum models fall below 50 percent directional accuracy. The practical upshot, if correct, is that simple tabular models outperform both deep learning and current quantum machine learning for this forecasting task.

What carries the argument

The mechanism that carries the argument is the benchmark itself: six models fit to the same one year of Curve Finance pool data under a common temporal split, evaluated by mean absolute error, root mean squared error, and directional accuracy. The quantum models additionally rely on quantum feature maps and parametrized ansatze (for the QNN) or a quantum kernel (for the QSVM) to encode the classical yield features into quantum states. The comparison protocol is what turns the metric differences into a general claim about model suitability.

What would settle it

Rerun the benchmark on the same 28 pools with a documented split date, equal per-model tuning budgets, and quantum circuits optimized against the validation objective; if the best quantum model reaches or exceeds XGBoost's 71.57 percent directional accuracy, the claim that quantum models underperform would collapse.

Watch

Extended reading notes

Core claim

The paper's central discovery is a clean ordering of six models on one shared benchmark: tree ensembles dominate, deep learning sits in between, and quantum models do no better than a coin toss. On the held-out test set, XGBoost achieves the best directional accuracy (71.57 percent) and Random Forest the lowest mean absolute error (1.77), while the LSTM, Transformer, quantum neural network, and quantum support vector machine with quantum feature maps all trail, with the two quantum models scoring below 50 percent directional accuracy. The author's claim is that this ordering reflects current model capability on real DeFi time series, not an artifact of a single favorable split.

Load-bearing premise

The load-bearing premise is that the comparison is fair and leakage-free: all six models were tuned with comparable effort, the temporal split keeps future information out of training, and the quantum feature maps and ansatze were chosen competently rather than as weak placeholders.

Editorial extensions

If this is right

  • A liquidity allocator choosing among these six methods should reach for XGBoost or Random Forest first, since they dominate on all three reported metrics.
  • Deep learning and quantum models do not currently offer an accuracy advantage on this data scale, so their higher complexity is not justified by predictive gain.
  • The reported numbers provide a reproducible baseline: any future model proposed for DeFi yield prediction can be compared against XGBoost's 71.57 percent directional accuracy and Random Forest's 1.77 test MAE on this dataset.
  • Because the quantum models sit below 50 percent directional accuracy, the paper implies that near-term quantum machine learning is not yet a viable tool for this class of financial time series.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the paper's comparison would be more persuasive if the tuning budgets and quantum circuit designs were reported; if the quantum models were given shallow encodings or fewer shots while the classical models received heavier hyperparameter tuning, the gap could narrow or invert.
  • Editorial inference: implicit in the results is that the predictive signal in Curve Finance yields is captured better by tabular feature interactions than by sequence structure; a direct test would be to feed the tree models lagged features versus the LSTM/Transformer inputs and compare margins.
  • Editorial inference: a concrete testable extension is to scale the quantum feature maps to more qubits or use hybrid training with a classical optimizer on the validation loss; if directional accuracy then crosses 50 percent and approaches the classical numbers, the 'quantum underperforms' conclusion would require qualification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper benchmarks six models (XGBoost, Random Forest, LSTM, Transformer, QNN, and QSVM-QNN) on one year of data from 28 Curve Finance pools, reporting test MAE, RMSE, and directional accuracy. The central claim is that classical tree ensembles outperform both deep learning and quantum models, with XGBoost achieving the highest directional accuracy (71.57%) and Random Forest the lowest MAE (1.77), while quantum models fall below 50% directional accuracy.

Significance. If the comparison is fair and leakage-free, the result is a practically useful benchmark for DeFi yield prediction, showing that current quantum machine learning is not competitive on this real-world time-series task. The paper explicitly aims to be reproducible, which is valuable. However, the abstract alone does not establish the statistical significance of the reported differences, does not include a trivial baseline, and provides no details on tuning effort or circuit design, all of which are essential for a benchmark to be credible. The headline conclusion is therefore plausible but not yet verified.

major comments (3)
  1. [Abstract (results paragraph)] No naive baseline such as persistence (lag-1) is reported. Without it, the 71.57% directional accuracy cannot be judged against a trivial predictor; if a persistence baseline achieves comparable or better accuracy, the practical advantage of the ensemble models would be much weaker. This is load-bearing for the claim that tree ensembles are 'robust' for this task.
  2. [Abstract (results paragraph)] No error bars, confidence intervals, or significance tests are provided. The difference between XGBoost (71.57%) and Random Forest (71.36%) is likely within noise, and even the gap between classical and quantum models may be dependent on a single train/test split. The abstract should report variance across seeds or time-series cross-validation, otherwise the ranking of models, especially among the classical methods, is not statistically supported.
  3. [Abstract (results paragraph)] The fairness of the quantum-classical comparison is unverifiable from the abstract. The abstract gives no information about the temporal split boundaries, hyperparameter search budgets for each model, the number of circuit layers, the feature-map choice, the optimizer, or the simulator/backend. The quantum models' directional accuracy below 50% suggests a systematic pipeline failure (e.g., an inexpressive encoding or poor convergence) rather than an inherent limitation of quantum methods. Without evidence that the quantum models received comparable tuning effort, the central conclusion that 'quantum models underperform' does not follow from the reported numbers.
minor comments (4)
  1. [Abstract (model list)] The acronym 'QSVM-QNN' is confusing: it appears to denote a quantum support vector machine with a quantum feature map, but the QNN suffix suggests a neural network. Please clarify the architecture and explain how it differs from the model labeled 'QNN'.
  2. [Abstract (first sentence)] The phrase 'growing demand for accurate yield and performance forecasting' is a generic motivation; it would be helpful to cite specific prior work on DeFi yield prediction to place this benchmark in context.
  3. [Abstract (reproducibility claim)] The abstract states 'This work offers a reproducible benchmark' but does not mention a code repository or data availability statement. Please include the repository URL or a statement about data access to support this claim.
  4. [Abstract (data description)] The sample size is not specified. 'One year of historical data from 28 Curve Finance pools' could be daily, hourly, or block-level data; the number of observations per pool directly affects the interpretability of the MAE/RMSE values and the complexity of the models. Please state the observation frequency and total sample size.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical benchmark with no derivation chain; no circularity identified from the abstract.

full rationale

This is an abstract-only review of an empirical benchmark paper. The claimed results (XGBoost and Random Forest outperform deep learning and quantum models on Curve Finance yield prediction) are obtained by fitting models to training data and evaluating them on test metrics such as MAE, RMSE, and directional accuracy. This is the standard non-circular evaluation setup: the models are not defined in terms of the test outcomes, and no fitted parameter is renamed as a prediction. There is no equation in the abstract that reduces to its own input, no self-citation invoked as a load-bearing theorem, and no ansatz smuggled in via citation. The only concerns are empirical fairness and reproducibility (e.g., comparable tuning budgets, leakage-free temporal split, competent quantum encodings), but those are correctness or verification risks, not circularity. Because no specific circular step can be quoted, the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No free constants, derived quantities, or invented entities appear in the abstract. The central claim rests on two empirical pillars that cannot be checked from the abstract: fair and leakage-free model comparison, and representativeness of the dataset. These are domain assumptions rather than mathematical axioms.

free parameters (2)
  • Model hyperparameters (six models) = undisclosed in abstract
    Each of XGBoost, Random Forest, LSTM, Transformer, QNN, and QSVM-QNN has tunable hyperparameters. Whether they were tuned with comparable budgets and without using the test set is not disclosed in the abstract, and the fairness of the headline comparison depends on this.
  • Feature set and preprocessing choices = undisclosed in abstract
    Which pool features feed the models, how they are scaled, and how the one year of data is split into train and test windows are not stated. These choices materially affect MAE, RMSE, and directional accuracy, and all six models share them.
assumptions (3)
  • domain assumption Temporal train/test split with no lookahead leakage
    The test MAE and directional accuracy numbers are meaningful only if test data did not influence training or hyperparameter selection. The abstract does not describe the split, so this standard ML hygiene assumption is unverified.
  • domain assumption One year of data from 28 pools is representative of DeFi yield dynamics
    The abstract generalizes from this single window and pool set to conclusions about quantum machine learning limitations on DeFi time series. Representatives is assumed rather than demonstrated.
  • domain assumption Quantum encodings and circuit designs are competently chosen
    Quantum model performance depends heavily on the feature map and ansatz. Without circuit details, sub-coin-flip directional accuracy for quantum models could reflect poor design rather than an inherent limitation of quantum machine learning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Classical and Quantum Models for DeFi Yield Prediction on Curve Finance." pith.science (2026). https://pith.science/paper/SG6M4M6V

@misc{pith2026250802685,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Classical and Quantum Models for DeFi Yield Prediction on Curve Finance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SG6M4M6V}},
  note         = {Machine review of arXiv:2508.02685}
}
read the original abstract

The rise of decentralized finance (DeFi) has created a growing demand for accurate yield and performance forecasting to guide liquidity allocation strategies. In this study, we benchmark six models, XGBoost, Random Forest, LSTM, Transformer, quantum neural networks (QNN), and quantum support vector machines with quantum feature maps (QSVM-QNN), on one year of historical data from 28 Curve Finance pools. We evaluate model performance on test MAE, RMSE, and directional accuracy. Our results show that classical ensemble models, particularly XGBoost and Random Forest, consistently outperform both deep learning and quantum models. XGBoost achieves the highest directional accuracy (71.57%) with a test MAE of 1.80, while Random Forest attains the lowest test MAE of 1.77 and 71.36% accuracy. In contrast, quantum models underperform with directional accuracy below 50% and higher errors, highlighting current limitations in applying quantum machine learning to real-world DeFi time series data. This work offers a reproducible benchmark and practical insights into model suitability for DeFi applications, emphasizing the robustness of classical methods over emerging quantum approaches in this domain.

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.