pith. sign in

arxiv: 2606.00060 · v1 · pith:UB6ERT4Xnew · submitted 2026-05-19 · 💱 q-fin.TR · cs.CE· cs.LG

Machine Learning-Based Bitcoin Trading Under Transaction Costs: Evidence From Walk-Forward Forecasting

Andrei Bysik , Robert \'Slepaczuk This is my paper

Pith reviewed 2026-06-30 17:37 UTC · model grok-4.3

classification 💱 q-fin.TR cs.CEcs.LG
keywords bitcoin tradingmachine learningtransaction costswalk-forward forecastingXGBoostcost-aware trading
0
0 comments X

The pith

A cost-aware execution filter turns machine learning forecasts of hourly Bitcoin returns into profitable strategies after ten basis point transaction costs.

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

The paper tests whether XGBoost, LSTM, and iTransformer forecasts of hourly BTC-USDT returns can produce positive net performance after costs in a 27-fold walk-forward setup on 2018-2026 data. Naive sign-based rules lose money once ten basis points in costs are applied, but a threshold that skips trades unless the forecast magnitude exceeds a cost-derived level sharply cuts turnover and restores gains. The strongest long-only XGBoost configuration then reaches annualised returns above 65 percent with a Sharpe ratio above one. Technical indicators sometimes add value while EGARCH features do not, and model or loss-function differences remain secondary. The central message is that the conversion of forecasts into trades, not raw predictability, is the binding constraint in this setting.

Core claim

A cost-aware execution filter that prevents trades unless the absolute forecast exceeds a transaction-cost-based threshold allows selected XGBoost long-only strategies to deliver annualised returns above 65 percent with Sharpe ratios above one on hourly BTC-USDT data, whereas naive sign-based strategies lose money under the same ten-basis-point costs.

What carries the argument

The cost-aware execution filter, which skips trades when forecast magnitude falls below a threshold calibrated to transaction costs, thereby reducing turnover.

If this is right

  • XGBoost produces descriptively stronger results than the tested neural networks under the same protocol.
  • Technical indicators can improve performance in selected configurations while EGARCH-derived features do not deliver uniform gains.
  • Loss-function choice and model-selection details have only secondary and statistically fragile effects.
  • The binding limit on hourly cryptocurrency trading performance is the translation of forecasts into position changes rather than forecast accuracy alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • High-frequency strategies in assets with persistent costs may benefit more from execution rules than from further gains in predictive power.
  • The filter approach could be tested on other liquid cryptocurrencies or equity futures to check whether the turnover-reduction effect generalises.
  • Variable or regime-dependent cost estimates might tighten or loosen the threshold and alter the reported Sharpe ratios.

Load-bearing premise

A fixed ten basis point transaction cost accurately represents real trading frictions and the chosen magnitude threshold is not overfitted to the 2018-2026 periods.

What would settle it

Re-running the identical 27-fold protocol on hourly BTC-USDT data after 2026 or with actual exchange fees plus slippage would show whether the reported returns persist.

Figures

Figures reproduced from arXiv: 2606.00060 by Andrei Bysik, Robert \'Slepaczuk.

Figure 1
Figure 1. Figure 1: compares the BTC/USDT price series with the corresponding hourly log-return series over the effective walk-forward evaluation window [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Walk-forward optimisation scheme Note: In each fold, the model is trained on 12 months of data, validated on the following 3 months, and tested on the next 3 months. The window advances by 3 months after each fold, producing 27 sequential test folds. The consolidated out-of-sample evaluation is the union of all held-out test windows. After the best configuration has been selected on the validation segment,… view at source ↗
Figure 3
Figure 3. Figure 3: Impact of transaction costs on naive strategy performance Note: Equity curves for long-only and long-short strategies. The left column shows results without transaction costs, while the right column reports results with transaction costs of 10 basis points per unit of turnover. Results use the OHLCV+TA+EGARCH feature set, MSE loss, and loss-best selection [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: compares baseline and cost-aware XGBoost strategies under the OHLCV+TA+EGARCH feature set, MSE loss, and three validation-set selection rules. The visual pattern is clear: baseline strategies deteriorate after costs, while the cost-aware versions generate substantially more stable equity curves [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of feature richness on cost-aware strategy performance Note: Equity curves for OHLCV, OHLCV+TA, and OHLCV+TA+EGARCH specifications under cost-aware execution with λ = 2.0. Results use XGBoost with MSE loss. The top row reports loss-based selection, while the bottom row reports mode-specific IR∗∗-based selection. The left column shows long-only strategies and the right column shows long-short strateg… view at source ↗
Figure 6
Figure 6. Figure 6: shows the cost-aware equity curves. In the long-only mode, all three architectures produce positive net performance after filtering, but XGBoost generates the strongest and most stable equity curve. In the long-short mode, XGBoost remains viable, while the neural-network strategies become difficult to interpret because cost-aware filtering suppresses almost all trades [PITH_FULL_IMAGE:figures/full_fig_p02… view at source ↗
Figure 7
Figure 7. Figure 7: compares the equity curves. In the long-only mode, MSE produces a visibly stronger equity path than MAE under cost-aware execution. In the long-short mode, MSE remains viable while MAE deteriorates [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: compares the equity curves across selectors. In the long-only mode, loss-best and IR∗∗-best produce similar trajectories, while IC-best is visibly weaker. In the long-short mode, IR∗∗-best is descriptively strongest [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sensitivity of strategy performance to the cost-aware threshold λ Note: The figure shows equity curves for different values of the cost-aware threshold λ. Lower values imply weaker filtering and more frequent trading, whereas higher values impose stricter execution thresholds and may suppress trading activity. The row λ = 0 corresponds to the unfiltered baseline strategy [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 10
Figure 10. Figure 10: and [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
read the original abstract

This paper investigates whether machine learning forecasts of hourly BTC-USDT returns can be converted into economically meaningful trading performance after transaction costs. Using approximately 70,000 hourly observations from 2018-2026, XGBoost, LSTM, and iTransformer are evaluated in a 27-fold walk-forward protocol. All three models produce positive gross trading performance in selected configurations, but naive sign-based strategies fail once transaction costs of ten basis points are imposed. A cost-aware execution filter, which prevents trades only when the forecast magnitude exceeds a transaction-cost-based threshold, sharply reduces turnover and restores profitability in selected configurations. The strongest long-only XGBoost strategy produces annualised returns above 65% with a Sharpe ratio above one. Additional tests show that technical indicators improve performance in selected cases, EGARCH-derived features do not provide uniformly robust gains, and XGBoost is descriptively stronger than the neural alternatives, although bootstrap evidence does not support formal statistical dominance. Loss-function and model-selection effects are secondary and statistically fragile. The results show that the main obstacle in hourly cryptocurrency trading is not only weak predictability, but also the way forecasts are converted into trades.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper evaluates whether ML forecasts (XGBoost, LSTM, iTransformer) of hourly BTC-USDT returns can yield economically meaningful trading profits after 10bp transaction costs. Using ~70k observations from 2018-2026 in a 27-fold walk-forward protocol, it reports that naive sign-based rules fail under costs but a cost-aware filter (trading only when |forecast| exceeds a transaction-cost-based threshold) reduces turnover and restores profitability in selected configurations. The strongest long-only XGBoost strategy achieves >65% annualized net returns with Sharpe >1. Additional analyses examine technical indicators, EGARCH features, model comparisons, and loss functions.

Significance. If the cost-aware threshold is a fixed, a-priori function of the 10bp cost alone and not tuned to performance, the work usefully demonstrates that execution rules, not just forecast accuracy, determine economic viability in high-frequency crypto trading. The 27-fold walk-forward design and multi-model comparison are strengths that allow assessment of robustness across regimes. The reported performance levels, however, are high enough to require unusually strong evidence of out-of-sample validity.

major comments (3)
  1. [Abstract and results] Abstract and results section: The headline claim of >65% annualized returns and Sharpe >1 for the long-only XGBoost strategy is presented without standard errors, bootstrap intervals, or p-values against a zero-return or buy-and-hold null. Given the 27-fold protocol and the emphasis on 'selected configurations,' this omission makes it impossible to judge whether the result is statistically distinguishable from noise or data-specific luck.
  2. [Methods (cost-aware filter)] Methods (cost-aware filter): The threshold is described as 'transaction-cost-based' and derived from the fixed 10bp cost and forecast magnitude. The manuscript does not state whether this exact mapping is held constant across all 27 folds or whether its level (or functional form) was chosen or refined by reference to realized net performance inside the walk-forward windows. If the latter, the filter introduces an extra degree of freedom that directly undermines the out-of-sample interpretation of the net-profitability result.
  3. [Results (model comparison)] Results (model comparison): The statement that 'XGBoost is descriptively stronger than the neural alternatives, although bootstrap evidence does not support formal statistical dominance' is load-bearing for the model-selection conclusion, yet the bootstrap procedure, number of replications, and exact test statistic are not detailed enough to verify that the lack of dominance is not simply low power.
minor comments (2)
  1. [Data section] The abstract states 'approximately 70,000 hourly observations' but the exact start/end dates, handling of missing hours, and any filtering for low-liquidity periods should be stated explicitly in the data section.
  2. [Methods] Notation for the forecast magnitude threshold and the precise functional mapping from 10bp cost to cutoff value should be given as an equation rather than described in prose.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our results and methods. We address each major comment below and indicate the revisions that will be incorporated.

read point-by-point responses

  1. Referee: [Abstract and results] Abstract and results section: The headline claim of >65% annualized returns and Sharpe >1 for the long-only XGBoost strategy is presented without standard errors, bootstrap intervals, or p-values against a zero-return or buy-and-hold null. Given the 27-fold protocol and the emphasis on 'selected configurations,' this omission makes it impossible to judge whether the result is statistically distinguishable from noise or data-specific luck.

    Authors: We agree that the absence of uncertainty measures limits the ability to assess whether the reported performance is distinguishable from noise. In the revised manuscript we will add block-bootstrap confidence intervals (resampling the 27 walk-forward periods) for annualized net returns and Sharpe ratios of the highlighted long-only XGBoost strategy, together with one-sided p-values against a zero-return null. These will be reported both in the abstract and in the main results tables. revision: yes

  2. Referee: [Methods (cost-aware filter)] Methods (cost-aware filter): The threshold is described as 'transaction-cost-based' and derived from the fixed 10bp cost and forecast magnitude. The manuscript does not state whether this exact mapping is held constant across all 27 folds or whether its level (or functional form) was chosen or refined by reference to realized net performance inside the walk-forward windows. If the latter, the filter introduces an extra degree of freedom that directly undermines the out-of-sample interpretation of the net-profitability result.

    Authors: The threshold is a fixed, a-priori function of the 10 bp cost and the absolute forecast value; its functional form and level are identical in every walk-forward fold and were never adjusted on the basis of realized net performance. We will add an explicit statement to this effect in the methods section, including the precise formula used, to eliminate any ambiguity regarding data snooping. revision: yes

  3. Referee: [Results (model comparison)] Results (model comparison): The statement that 'XGBoost is descriptively stronger than the neural alternatives, although bootstrap evidence does not support formal statistical dominance' is load-bearing for the model-selection conclusion, yet the bootstrap procedure, number of replications, and exact test statistic are not detailed enough to verify that the lack of dominance is not simply low power.

    Authors: We accept that the bootstrap comparison requires fuller documentation. The revision will specify that we employ a block bootstrap with 1,000 replications, resampling entire walk-forward periods to preserve temporal dependence, and test the null that the difference in Sharpe ratios (or annualized returns) equals zero. We will also report the resulting p-values and note the power implications of the modest number of folds. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper applies standard walk-forward validation across 27 folds on the 2018-2026 sample and converts ML forecasts into trades via a cost-aware filter whose threshold is described as a direct function of the fixed 10 bp transaction cost and forecast magnitude. No equations or text indicate that this threshold (or any other hyperparameter) is fitted or refined against realized net performance within or across folds; the filter is presented as an a-priori economic rule rather than a data-driven prediction. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked to justify core choices. The reported >65% annualized returns and Sharpe >1 are empirical outputs of the protocol, not inputs redefined by construction. The derivation chain from forecasts to net performance is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on a data-tuned execution threshold and standard assumptions about market efficiency and data stationarity in the walk-forward setup; no new entities are introduced.

free parameters (1)
  • cost-aware threshold = derived from 10bp
    The magnitude threshold for executing trades is set using the 10bp transaction cost and is likely optimized on the evaluation data.
axioms (1)
  • domain assumption Walk-forward validation with 27 folds yields unbiased estimates of live trading performance
    Invoked in the evaluation protocol description

pith-pipeline@v0.9.1-grok · 5736 in / 1306 out tokens · 42308 ms · 2026-06-30T17:37:34.075687+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

  1. [1]

    Complexity 2018, 8983590

    Anticipating Cryptocurrency Prices Using Machine Learning. Complexity 2018, 8983590. doi:10.1155/2018/8983590. Asness, C.S., Moskowitz, T.J., Pedersen, L.H.,

    work page doi:10.1155/2018/8983590 2018

  2. [2]

    The Journal of Finance 68, 929–985

    Value and momentum everywhere. The Journal of Finance 68, 929–985. doi:10.1111/jofi.12021. Bailey, D.H., Borwein, J.M., López de Prado, M., Zhu, Q.J.,

    work page doi:10.1111/jofi.12021

  3. [3]

    The probability of backtest overfitting.The Journal of Computational Finance, 20(4):39–69, 2017

    The probability of backtest overfitting. The Journal of Computational Finance 20, 39–69. doi:10.21314/JCF.2016.322. Bieganowski, B., Ślepaczuk, R.,

    work page doi:10.21314/jcf.2016.322 2016

  4. [4]

    Binance,

    doi:10.1186/s40537-025-01267-7. Binance,

    work page doi:10.1186/s40537-025-01267-7

  5. [5]

    https://www.binance.com/en/fee/futureFee

    USD-MFuturesTradingFeeRate. https://www.binance.com/en/fee/futureFee. Accessed: 2026-04-30. Campbell, J.Y., Lo, A.W., MacKinlay, A.C.,

    2026

  6. [6]

    XGBoost: A scalable tree boosting system, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM. pp. 785–794. doi:10.1145/2939672.2939785. Cohen, G.,

    work page doi:10.1145/2939672.2939785

  7. [7]

    Review of Quantitative Finance and Accounting 61, 395–409

    Intraday algorithmic trading strategies for cryptocurrencies. Review of Quantitative Finance and Accounting 61, 395–409. doi:10.1007/s11156-023-01139-2. DeMiguel, V., Garlappi, L., Uppal, R.,

    work page doi:10.1007/s11156-023-01139-2

  8. [8]

    Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? Review of Financial Studies, 22 0 (5): 0 1915--1953, 2009

    Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? The Review of Financial Studies 22, 1915–1953. doi:10.1093/rfs/hhm075. Fama, E.F.,

    work page doi:10.1093/rfs/hhm075 1915

  9. [9]

    The Journal of Finance 25, 383–417

    Efficient capital markets: A review of theory and empirical work. The Journal of Finance 25, 383–417. doi:10.2307/2325486. Grinsztajn, L., Oyallon, E., Varoquaux, G.,

    work page doi:10.2307/2325486

  10. [10]

    Why do tree-based models still outperform deep learning on typical tabular data?, in: Advances in Neural Information Processing Systems, pp. 507–520. doi:10.48550/arXiv.2207.08815. Grudniewicz, J., Ślepaczuk, R.,

    work page doi:10.48550/arxiv.2207.08815

  11. [11]

    Experimental searches for the chiral magnetic effect in heavy-ion collisions.Prog

    Application of machine learning in algorithmic investment strategies on global stock markets. Research in International Business and Finance 66, 102052. doi:10.1016/j. ribaf.2023.102052. Gu, S., Kelly, B., Xiu, D.,

    work page doi:10.1016/j 2023

  12. [12]

    Moritz Hardt and Celestine Mendler-Dünner

    Empirical asset pricing via machine learning. The Review of Financial Studies 33, 2223–2273. doi:10.1093/rfs/hhaa009. Hochreiter, S., Schmidhuber, J.,

    work page doi:10.1093/rfs/hhaa009

  13. [13]

    Long Short-Term Memory , year =

    Long short-term memory. Neural Computation 9, 1735–1780. doi:10.1162/neco.1997.9.8.1735. Jegadeesh, N., Titman, S.,

    work page doi:10.1162/neco.1997.9.8.1735 1997

  14. [14]

    The Journal of Finance 48, 65–91

    Returns to buying winners and selling losers: Implications for stock market efficiency. The Journal of Finance 48, 65–91. doi:10.1111/j.1540-6261.1993.tb04702.x. Kashif, K., Ślepaczuk, R.,

    work page doi:10.1111/j.1540-6261.1993.tb04702.x 1993

  15. [15]

    Knowledge-Based Systems 320, 113563

    LSTM-ARIMA as a hybrid approach in algorithmic investment strategies. Knowledge-Based Systems 320, 113563. doi:10.1016/j.knosys.2025.113563. Kryńska, K., Ślepaczuk, R.,

    work page doi:10.1016/j.knosys.2025.113563 2025

  16. [16]

    doi:10.2139/ssrn.4628806

    Daily and intraday application of various architectures of the LSTM model in algorithmic investment strategies on Bitcoin and the S&P 500 index. doi:10.2139/ssrn.4628806. Künsch, H.R.,

    work page doi:10.2139/ssrn.4628806

  17. [17]

    The Annals of Statistics 17, 1217–1241

    The Jackknife and the Bootstrap for General Stationary Observations. The Annals of Statistics 17, 1217–1241. doi:10.1214/aos/1176347265. Liaras, E., Nerantzidis, M., Alexandridis, A.,

    work page doi:10.1214/aos/1176347265

  18. [18]

    Michańków, J., Sakowski, P., Ślepaczuk, R., 2024a

    doi:10.3390/s22030917. Michańków, J., Sakowski, P., Ślepaczuk, R., 2024a. Generalized mean absolute directional loss as a solution to overfitting and high transaction costs in machine learning models used in high-frequency algorithmic investment strategies. doi:10.48550/arXiv.2412.18405. Michańków, J., Sakowski, P., Ślepaczuk, R., 2024b. Mean absolute dir...

    work page doi:10.3390/s22030917 2024

  19. [19]

    chapter 6, pp. 328–383. doi:10.1016/B978-0-444-53683-9.00006-2. Sebastião, H., Godinho, P.,

    work page doi:10.1016/b978-0-444-53683-9.00006-2

  20. [20]

    Stefaniuk, F., Ślepaczuk, R.,

    doi:10.1186/s40854-020-00217-x. Stefaniuk, F., Ślepaczuk, R.,

    work page doi:10.1186/s40854-020-00217-x

  21. [21]

    Expert Systems with Applications 317, 131599

    A neural network informer in algorithmic investment strategies on high- frequency bitcoin data. Expert Systems with Applications 317, 131599. doi:10.1016/j.eswa.2026.131599. 42

    work page doi:10.1016/j.eswa.2026.131599 2026