Paper: arXiv 2512.22476

Authors: Kaihong Deng

Abstract

Backtests of cryptocurrency perpetual futures are fragile when they ignore microstructure frictions and reuse evaluation windows during parameter search. We study four liquid perpetuals (BTC/USDT, ETH/USDT, SOL/USDT, AVAX/USDT) and quantify how execution delay, funding, fees, and slippage can inflate reported performance. We introduce AutoQuant, an execution-centric, alpha-agnostic framework for auditable strategy configuration selection. AutoQuant encodes strict T+1 execution semantics and no-look-ahead funding alignment, runs Bayesian optimization under realistic costs, and applies a two-stage double-screening protocol across held-out rolling windows and a cost-sensitivity grid. We show that fee-only and zero-cost backtests can materially overestimate annualized returns relative to a fully costed configuration, and that double screening tends to reduce drawdowns under the same strict semantics even when returns are not higher. A CSCV/PBO diagnostic indicates substantial residual overfitting risk, motivating AutoQuant as validation and governance infrastructure rather than a claim of persistent alpha. Returns are reported for small-account simulations with linear trading costs and without market impact or capacity modeling.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 5.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced Bayesian optimization and statistical diagnostics (CSCV/PBO) for rigorous validation, but its empirical testing is constrained to a limited set of assets and small-account simulations without market impact modeling.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Quantify backtest fragility & design auditable tuning"] --> B{"Inputs: 4 Crypto Perpetuals"}
  B --> C["Methodology: AutoQuant Framework"]
  C --> D["Execution: Bayesian Optimization<br>with T+1 semantics & costs"]
  D --> E{"Double Screening:<br>Rolling Windows + Cost Sensitivity"}
  E --> F["Outcomes: Overfitting Diagnostics<br>and Validated Configs"]
  F --> G["Key Findings:<br>Costs reduce returns; DCS lowers drawdowns"]

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 proposes AutoQuant, an execution‑centric, alpha‑agnostic framework for auditable strategy configuration selection in cryptocurrency perpetual futures.

Source sentence
We propose AutoQuant, an execution-centric and alpha-agnostic framework that can be viewed as an auditable expert system for strategy configuration selection.

Assumes 6

  • Strict T+1 execution semantics are enforced, with funding applied only after the bar closes. execution
    Source sentence
    Funding is charged on the signed nominal exposure using an 8-hour funding rate series, scaled to the bar length.
  • Pure taker execution is used (maker ratio = 0), so fees are taker fees only. market
    Source sentence
    our STRICT4H case study uses pure taker execution (maker ratio = 0), so feff reduces to the taker fee.
  • Trading costs are modeled conservatively and applied uniformly across backtesting and simulation. model
    Source sentence
    A cornerstone of our framework is a realistic trading-cost model, which aims to provide a comprehensive and conservative estimate of market frictions. The model is applied uniformly across both backtesting and simulation stages to ensure consistency.
  • No market impact or capacity constraints are modeled; results are for small‑account simulations. other
    Source sentence
    Returns are reported for small-account simulations with linear trading costs and without market impact or capacity modeling.
Show 2 more
  • Funding rates are aligned as a step function carried forward between 8‑hour timestamps. data
    Source sentence
    dataset includes a synchronized time series of funding rates extracted at 8-hour intervals, which we align to the 4-hour time axis as a step function (carried forward between funding timestamps under a carry-forward alignment).
  • The cost profile assumes an initial equity of 10,000 USDT, max leverage 5×, notional cap 50,000 USDT. model
    Source sentence
    In our BTC/USDT STRICT4H experiments, the baseline cost profile assumes an initial equity of 10,000 USDT, a maximum leverage of 5×, a notional cap of 50,000 USDT, a taker fee of 4 bps per trade, and a base slippage of 2 bps per trade.

Data

Backtest sample: 2021-12 – 2025-10 · BTC/USDT · crypto · intraday

  • BTC/USDT perpetual futures 4‑hour OHLCV bars · Binance, Bybit, OKX market-prices public
    2019-09-08 to 2021-12-31 (core) and exte · BTC/USDT perpetual contract
    Source sentence
    we utilize a comprehensive dataset of 4-hour (4h) OHLCV bars sourced from major centralized exchanges, with prices and volumes constructed from exchange-level perpetual futures markets
  • ETH/USDT, SOL/USDT, AVAX/USDT perpetual futures 4‑hour OHLCV bars · major centralized exchanges market-prices public
    2019-09-08 to 2021-12-31 (core) and exte · ETH/USDT, SOL/USDT, AVAX/USDT perpetual contracts
    Source sentence
    we report parallel replications on ETH/USDT (large-cap) and SOL/USDT and AVAX/USDT (mid-cap) perpetual contracts using the same STRICT4H family and evaluation scripts.
  • 8‑hour funding rate series · exchange‑provided funding data alternative public
    2019-09-08 to 2025-10-14 · same four perpetual contracts
    Source sentence
    our dataset includes a synchronized time series of funding rates extracted at 8-hour intervals, which we align to the 4-hour time axis as a step function

Recognised sources: Binance Bybit OKX

Costs

Modeled Costs are part of the reported headline results.

Taker fee 4 bps per trade, base slippage 2 bps per trade, fallback funding 1 bps per 8‑hour period; cost model applied uniformly in backtest and simulation.

Source sentence
... a taker fee of 4 bps per trade, and a base slippage of 2 bps per trade. The cost profile also includes a fallback funding rate of 1 bps per 8-hour period

Finds 5

Reported performance (AutoQuant two-stage selection (STRICT4H)): Sharpe 1.67 · annual return 105.8% · max drawdown 20.9%

  • Zero‑cost and fee‑only backtests materially overestimate annualized returns compared to fully costed configurations. performance
    Source sentence
    We show that fee-only and zero-cost backtests can materially overestimate annualized returns relative to a fully costed configuration
  • The two‑stage double‑screening protocol reduces drawdowns even when returns are not higher. performance
    Source sentence
    double screening tends to reduce drawdowns under the same strict semantics even when returns are not higher
  • Cost‑sensitivity analysis shows that moderate underestimation of frictions can compress annualized returns and Sharpe ratios. performance
    Source sentence
    Across plausible cost scenarios, annualized returns and Sharpe can compress materially relative to the baseline, underscoring that moderate underestimation of trading frictions can distort conclusions about apparent profitability
  • Drawdown‑bucket risk‑budgeting variants improve annualized return by 24–33 pp and reduce max drawdown by 14–26 pp versus the baseline. performance
    Source sentence
    drawdown‑bucket risk‑budgeting variants improve annualized return by roughly 24–33 percentage points and reduce maximum drawdown by about 14–26 percentage points relative to the STRICT4H baseline
  • Deflated Sharpe‑ratio (DSR) diagnostic indicates substantial residual overfitting risk, suggesting the framework is for validation rather than claiming persistent alpha. statistical
    Source sentence
    A CSCV/PBO diagnostic indicates substantial residual overfitting risk, motivating AutoQuant as validation and governance infrastructure rather than a claim of persistent alpha

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 42 related papers). Coverage is partial, so this is neither replication nor refutation.

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 (46)

  1. The Limits of Arbitrage (1997) — link
  2. The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality (2014) — link
  3. Advances in Financial Machine Learning (2018) — link
  4. Seven Sins of Quantitative Investing (2019)
  5. Practical Bayesian Optimization of Machine Learning Algorithms (2012) — link
  6. Algorithms for Hyper-Parameter Optimization (2011) — link
  7. Random Search for Hyper-Parameter Optimization (2012) — link
  8. Differential Evolution – A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces (1997) — link
and 38 more
  1. The CMA Evolution Strategy: A Comparing Review (2006) — link
  2. Genetic Algorithms in Search, Optimization, and Machine Learning (1989)
  3. backtrader: Python Backtesting Library for Trading Strategies (2025)
  4. The Jackknife and the Bootstrap for General Stationary Observations (1989) — link
  5. Zipline: A Pythonic Algorithmic Trading Library (2025)
  6. Freqtrade: Free, Open Source Crypto Trading Bot (2025)
  7. FinRL: Deep Reinforcement Learning Framework for Quantitative Finance (2025)
  8. Trading and Arbitrage in Cryptocurrency Markets (2020) — link
  9. From Mining to Markets: The Evolution of Bitcoin Transaction Fees (2019)
  10. Decentralized Finance: On Blockchain- and Smart Contract-Based Financial Markets (2021) — link
  11. Automated Machine Learning: Methods, Systems, Challenges (2019) — link
  12. Optuna: A Next-Generation Hyperparameter Optimization Framework (2019) — link
  13. The Probability of Backtest Overfitting (2017) — link
  14. Algorithmic and High-Frequency Trading (2015)
  15. The Science of Algorithmic Trading and Portfolio Management (2014)
  16. Empirical Asset Pricing via Machine Learning (2020)
  17. Deep Learning in Asset Pricing (2021)
  18. Deep Learning with Long Short-Term Memory Networks for Financial Market Predictions (2018)
  19. Learning to Trade via Direct Reinforcement (2001)
  20. Optimal Execution of Portfolio Transactions (2001) — link
  21. No-dynamic-arbitrage and market impact (2010) — link
  22. Anomalous Price Impact and the Critical Nature of Liquidity in Financial Markets (2011) — link
  23. Data-Snooping, Technical Trading Rule Performance, and the Bootstrap (1999) — link
  24. A Reality Check for Data Snooping (2000) — link
  25. A Test for Superior Predictive Ability (2005) — link
  26. \ldots and the Cross-Section of Expected Returns (2015) — link
  27. The FinTech Opportunity (2016) — link
  28. Digital Finance and FinTech: current research and future research directions (2017) — link
  29. Risks and Returns of Cryptocurrency (2020) — link
  30. Common Risk Factors in Cryptocurrency (2022) — link
  31. Perpetual Futures Pricing — → 2310.11771
  32. Pricing Perpetual Futures (2023) — link
  33. Perpetual Futures and Basis Risk: Evidence from Cryptocurrency (2025) — link
  34. Perpetual Futures Contracts and Cryptocurrency Market Quality (2025) — link
  35. Perpetual future contracts in centralized and decentralized exchanges: Mechanism and traders’ behavior (2024) — link
  36. Trading Activity and Price Discovery in Bitcoin Futures Markets (2021) — link
  37. Impact of Bitcoin Futures on the Informational Efficiency of Bitcoin Spot Market (2020) — link
  38. A comparison of machine learning and econometric models for pricing perpetual Bitcoin futures and their application to algorithmic trading (2023) — link

1 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.