Paper: arXiv 2508.02685
Authors: Chi-Sheng Chen, Aidan Hung-Wen Tsai
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.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper integrates advanced machine learning models including quantum neural networks with specific circuit architectures and hybrid feature maps, indicating moderate-to-high mathematical sophistication. It demonstrates strong empirical rigor through a detailed benchmarking study using real DeFi data (Curve Finance pools), clear train/test splits, feature engineering, and standard evaluation metrics (MAE, RMSE, directional accuracy).
Research Flowchart
flowchart TD
A["Research Goal<br>Benchmark classical vs. quantum models<br>for DeFi yield prediction on Curve Finance"] --> B["Data: 1 Year Historical Data<br>28 Curve Finance Pools"]
B --> C["Methodology: Model Benchmarking<br>Train/Test Split & Evaluation Metrics"]
C --> D{"Computational Process"}
D --> E["Classical Models<br>XGBoost, Random Forest, LSTM, Transformer"]
D --> F["Quantum Models<br>QNN, QSVM-QNN"]
E & F --> G["Key Findings & Outcomes"]
G --> H["Classical Ensemble Models Win<br>XGBoost MAE: 1.80 | Acc: 71.57%<br>Random Forest MAE: 1.77 | Acc: 71.36%"]
G --> I["Quantum Models Underperform<br>Accuracy < 50% | Higher Errors<br>Current limitations in DeFi application"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper presents a comprehensive benchmark of six machine learning models on a newly curated dataset from 28 Curve Finance pools over a full year.
Source sentence
In this work, we present a comprehensive benchmark study of six machine learning models on a newly curated dataset derived from 28 Curve Finance
Assumes 3
- Models are trained and evaluated using identical feature sets and an 80/20 chronological data split. data
Source sentence
Under identical feature sets and an 80/20 chronological split, Random Forest achieved the lowest test MAE (1.77), while XGBoost delivered the highest directional accuracy (71.57%).
- The time series data is temporally complete with no missing timestamps on the six‑hour grid. data
Source sentence
The series is temporally complete with no missing timestamps on the six-hour grid, ensuring a clean foundation for downstream forecasting tasks.
- A six‑hour sampling frequency is sufficient to capture intra‑day fluctuations typical of DeFi markets. market
Source sentence
This sampling frequency is sufficient to capture intra-day fluctuations typical of DeFi markets while maintaining an almost year-long horizon.
Data
- Curve Finance pools dataset crypto-onchain unknown
Source sentence
We compiled historical time series from 28 Curve Finance liquidity pools, covering the period from 2024-07-20 15:31:05 to 2025-07-20 09:31:05, for a total span of 364 days...
Costs
Not applicable
Finds 5
- XGBoost achieved the highest directional accuracy (71.57%) with a test MAE of 1.80. performance
Source sentence
the highest directional accuracy (71.57%) with a test MAE of 1.80
- Random Forest attained the lowest test MAE (1.77) and a directional accuracy of 71.36%. performance
Source sentence
Random Forest attains the lowest test MAE of 1.77 and 71.36% accuracy
- Quantum models (QNN and QSVM‑QNN) had directional accuracy below 50% and higher test MAE (2.25–2.33). negative
Source sentence
Quantum models—QNN and QSVM‑QNN—performed worse than classical baselines, with higher test MAE (2.25–2.33) and lower accuracy (<50%).
- Deep learning models (LSTM and Transformer) lagged behind classical models with directional accuracy around 49–51% and higher errors. performance
Source sentence
Among deep learning models, both LSTM and Transformer lagged behind with directional accuracy around 49–51%.
- XGBoost showed near‑perfect training accuracy (99.85%) suggesting possible overfitting, though test performance remained strong. statistical
Source sentence
While XGBoost achieved nearly perfect training accuracy (99.85%), its test MAE and accuracy are close to Random Forest, suggesting possible overfitting.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 33 related papers). Coverage is partial, so this is neither replication nor refutation.
3 papers in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (30)
- Quantum adaptive self-attention for quantum transformer models (2025) — link
- Quantum-enhanced channel mixing in rwkv models for time series forecasting (2025) — link
- Quantum long short-term memory (2022)
- Learning to program variational quantum circuits with fast weights (2024)
- Long short-term memory (1997)
- Attention is all you need (2017)
- A multi-source dataset of urban life in the city of Milan and the Province of Trentino (2015)
- Long Short-Term Memory (1997) — link
and 22 more
- Quantum Machine Learning (2017) — link
- Quantum Long Short-Term Memory (2020) — link
- RWKV: Reinventing RNNs for the Transformer Era (2023)
- Learning to Program Variational Quantum Circuits with Fast Weights (2024) — link
- Time series analysis: forecasting and control (1970)
- Forecasting sales by exponentially weighted moving averages (1960)
- Forecasting seasonals and trends by exponentially weighted moving averages (1957)
- Forecasting with exponential smoothing: the state space approach (2008)
- Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021) — link
- Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021) — link
- Dynamic population mapping using mobile phone data (2014)
- Urban informatics and future cities (2015)
- Quantum neural networks with data re-uploading for urban traffic time series forecasting (2025) — link
- Quantum long short-term memory (QLSTM) vs. classical LSTM in time series forecasting: a comparative study in solar power forecasting (2024)
- The quest for a quantum neural network (2014)
- Xgboost: A scalable tree boosting system (2016)
- Quantum support vector machine for big data classification (2014)
- Random forests (2001)
- Numhtml: Numeric-oriented hierarchical transformer model for multi-task financial forecasting (2022)
- Quantum finance (2002)
- Quantum finance: Exploring the implications of quantum computing on financial models (2025)
- Curve Stablecoin Whitepaper (2023)
0 of 30 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.