Paper: arXiv 2512.04099

Authors: Andrzej Tokajuk, Jarosław A. Chudziak

Abstract

Forecasting cryptocurrency prices is hindered by extreme volatility and a methodological dilemma between information-scarce univariate models and noise-prone full-multivariate models. This paper investigates a partial-multivariate approach to balance this trade-off, hypothesizing that a strategic subset of features offers superior predictive power. We apply the Partial-Multivariate Transformer (PMformer) to forecast daily returns for BTCUSDT and ETHUSDT, benchmarking it against eleven classical and deep learning models. Our empirical results yield two primary contributions. First, we demonstrate that the partial-multivariate strategy achieves significant statistical accuracy, effectively balancing informative signals with noise. Second, we experiment and discuss an observable disconnect between this statistical performance and practical trading utility; lower prediction error did not consistently translate to higher financial returns in simulations. This finding challenges the reliance on traditional error metrics and highlights the need to develop evaluation criteria more aligned with real-world financial objectives.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: reported Sharpe ratio of 3.83 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: The paper employs advanced deep learning mathematics (dual-attention transformers, tokenization, embeddings, MSE optimization) and demonstrates high empirical rigor through extensive benchmarking against 11 models, hyperparameter tuning, and explicit discussion of trading utility vs. statistical metrics.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Balance information-scarce vs.<br>noise-prone forecasting in<br>cryptocurrency time series"] --> B["Methodology: Partial-Multivariate Transformer<br>(PMformer) with<br>strategic feature subset selection"]
  B --> C["Data & Inputs:<br>BTCUSDT & ETHUSDT<br>Daily Returns"]
  C --> D["Computational Process:<br>Compare PMformer vs.<br>11 benchmarks (Classical &<br>Deep Learning models)"]
  D --> E{"Key Findings & Outcomes"}
  E --> F["1. Statistical Performance:<br>PMformer achieves significant<br>accuracy improvement"]
  E --> G["2. Practical Utility:<br>Disconnect observed:<br>Low error ≠ High trading returns"]
  E --> H["Implication:<br>Need for evaluation criteria<br>aligned with real-world<br>financial objectives"]

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 the Partial-Multivariate Transformer (PMformer) for cryptocurrency price forecasting, balancing information scarcity and noise.

Source sentence
We apply the Partial-Multivariate Transformer (PMformer) to forecast daily returns for BTCUSDT and ETHUSDT, benchmarking it against eleven classical and deep learning models.

Assumes 6

  • The prediction target is the next‑day logarithmic return. data
    Source sentence
    The prediction target is the next-day logarithmic return, rt = ln(Pt/Pt−1), a standard variance-stabilizing transform for financial series [22].
  • All numeric features are scaled to [0,1] by Min–Max normalization fitted on the training split. data
    Source sentence
    All numeric features are scaled to [0, 1] by Min–Max normalization [25]; scaling parameters are fit on the training split and applied to the validation/test set to prevent leakage.
  • The dataset is split chronologically into 70% train, 20% validation, and 10% test. data
    Source sentence
    The dataset is split chronologically into 70% train, 20% validation, and 10% test.
  • Trading follows a simple sign‑based rule: long if predicted return > 0, short otherwise. execution
    Source sentence
    The trading strategy is a simple sign-based rule: go long if predicted return ˆrt+1 > 0 and short otherwise.
Show 2 more
  • PMformer trains on sampled feature subsets, acting as an implicit adaptive feature selection step that curbs variance. model
    Source sentence
    Because PMformer trains on sampled feature subsets, this acts as an implicit, adaptive feature selection step that curbs variance without exhaustive multivariate inputs [19].
  • Technical indicators (SMA, EMA, RSI, CCI, ATR, MACD) capture trend, momentum, and volatility sufficiently for the partial‑multivariate approach. model
    Source sentence
    This compact set captures trend, momentum, and volatility while remaining compatible with partial-multivariate training.

Data

Backtest sample: BTC/USDT · crypto · daily

  • BTCUSDT and ETHUSDT daily data · Binance market-prices public
    Oct. 5, 2017 to May 20, 2025 · Bitcoin (BTCUSDT) and Ethereum (ETHUSDT)
    Source sentence
    We use daily Bitcoin (BTCUSDT) and Ethereum (ETHUSDT) data from Binance from Oct. 5, 2017 to May 20, 2025.

Recognised sources: Binance

Costs

Not modeled A trading or execution result is reported without costs.

Finds 5

Reported performance (Partial-Multivariate Transformer (PMformer)): Sharpe 3.83 · total return 20.6% · max drawdown 13.8%

  • PMformer achieved the lowest MSE on BTCUSDT (6.5798 × 10⁻⁴), confirming its superior statistical accuracy. predictive-accuracy
    Source sentence
    PMformer achieved the highest statistical accuracy, posting the lowest MSE (6.5798 × 10−4) among all tested models.
  • On BTCUSDT, PMformer attained the highest directional accuracy (59.8%) and the lowest maximum drawdown (‑13.8%). performance
    Source sentence
    PMformer provided superior risk control, securing the lowest maximum drawdown (-13.8%) and the highest directional accuracy (59.8%).
  • FEDformer yielded a higher total ROI (+38.1%) and Sharpe Ratio (4.54) than PMformer, highlighting a trade‑off between return and risk. performance
    Source sentence
    While FEDformer yielded a higher Total ROI (+38.1%) and Sharpe Ratio (4.54), PMformer provided superior risk control...
  • For ETHUSDT, PMformer had the lowest forecasting errors but all models, including PMformer, generated negative total ROI, indicating prediction accuracy did not translate to profit. negative
    Source sentence
    PMformer confirmed expectations on the forecasting front (yielding the lowest errors for ETHUSDT as expected), it also provided a cautionary confirmation that prediction accuracy alone is not enough for successful trading...
  • The study reveals a pronounced disconnect between statistical accuracy and practical trading utility across both assets. statistical
    Source sentence
    Our analysis reveals a pronounced disconnect between statistical accuracy and practical trading utility. For both assets, the models with the lowest prediction errors were frequently not the most profitable.

Checked by others

Supported by

  • When Forecast Accuracy Fails: Rank Correlation and Decision Quality in Multi-Market Battery Storage Optimization (2026)
    On the question: Do lower forecasting errors guarantee higher economic profit in market decisions?
    Compare the findings

    This paper: For ETHUSDT, PMformer had the lowest forecasting errors but all models, including PMformer, generated negative total ROI, indicating prediction accuracy did not translate to profit.

    That paper: Persistence forecasts capture only 32.8% of oracle value despite high MAE, whereas ML‑based forecasts with τ≈0.9 capture >99% of oracle value, showing misalignment between MAE and economic performance.

    Both studies show error metrics mislead profit: A's PMformer has lowest errors yet negative ROI; B's persistence MAE captures only 32.8% oracle value, while ML τ≈0.9 captures >99% despite MAE differences.

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 (33)

  1. A critical analysis of volatility surprise in Bitcoin cryptocurrency and other financial assets (2021)
  2. An introduction to variable and feature selection (2003)
  3. Volatility estimation for Bitcoin: A comparison of GARCH models (2017)
  4. Wrappers for feature subset selection (1997)
  5. Computational methods of feature selection (2007)
  6. Partial-Multivariate Model for Forecasting (2024) — link
  7. Quantifying Cryptocurrency Unpredictability: A Comprehensive Study of Complexity and Forecasting — → 2502.09079
  8. How well do investor sentiment and ensemble learning predict Bitcoin prices? (2023)
and 25 more
  1. Regression shrinkage and selection via the lasso (1996)
  2. The inefficiency of Bitcoin (2016)
  3. Attention is all you need (2017)
  4. Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
  5. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2021)
  6. itransformer: Inverted transformers are effective for time series forecasting (2023) — link
  7. Long short-term memory (1997)
  8. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting (2022)
  9. A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
  10. PMformer: A novel informer-based model for accurate long-term time series prediction (2025)
  11. Time series analysis: forecasting and control (2015)
  12. Are transformers effective for time series forecasting? (2023)
  13. Forecasting stock market prices using machine learning and deep learning models: A systematic review, performance analysis and discussion of implications (2023)
  14. Univariate and multivariate machine learning forecasting models on the price returns of cryptocurrencies (2021)
  15. Taming the curse of dimensionality: quantitative economics with deep learning (2024)
  16. Feature selection for classification (1997)
  17. Analysis of financial time series (2005)
  18. Smarter trading (1995)
  19. Technical analysis of the financial markets: A comprehensive guide to trading methods and applications (1999)
  20. Deep learning (2016)
  21. Applying Informer for Option Pricing: A Transformer-Based Approach — → 2506.05565
  22. ElliottAgents: A natural language-driven multi-agent system for stock market analysis and prediction (2024)
  23. Toward predictive stock trading with hidformer integrated into reinforcement learning strategy (2024)
  24. Algorithmic trading: winning strategies and their rationale (2013)
  25. Algorithmic trading & DMA : an introduction to direct access trading strategies (2010)

2 of 33 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.