Does complexity help forecast Bitcoin?

The question

Does combining a more complex model with sentiment signals improve Bitcoin price forecasting?

In the one-day Bitcoin test, the model does not outperform persistence.

Persistence forecasts the last observed price. We study whether a Transformer, which learns relationships between inputs, improves on that reference when it incorporates sentiment at multiple time scales.

Data and comparison

We use daily price and volume data for Bitcoin, Ethereum and BNB from Binance and Coinbase, together with the Fear and Greed Index from Alternative.me. Bitcoin is the main target. Wavelets separate sentiment into faster and slower components within historical windows; the Transformer combines them with technical indicators.

The paper describes a temporal training, validation and test split and compares the model with persistence and other alternatives. We refer to the task here as price forecasting; the paper gives the details of the method and evaluation (pages 6–14).

The measured error

RMSE summarises the distance between forecasts and observed prices, penalising large errors more heavily. Lower values mean less error. Table 2, page 14, reports USD 2005 for Wavelet–Transformer and USD 1986 for persistence.

The model has 2.1 million parameters, the values adjusted during training (Table 6, page 18). That complexity does not improve forecasting in this test.

Consult the tables and method in the paper.

Scope and limitations

• The result concerns this model and this test. It does not show that every model fails or that every market is unpredictable.
• The study measures forecasting errors. It does not evaluate agents, trades, fees or slippage.
• A complete reproduction has not been verified. This page summarises the published paper.

Paper and citation

Aldan Jay; Rafael Berlanga. (2026). Algorithmic Complexity vs. Market Efficiency: Evaluating Wavelet–Transformer Architectures for Cryptocurrency Price Forecasting. Algorithms · 19(2) · 101.

DOI: 10.3390/a19020101 · CC BY 4.0

@article{jay2026algorithmic,
  title = {Algorithmic Complexity vs. Market Efficiency: Evaluating Wavelet–Transformer Architectures for Cryptocurrency Price Forecasting},
  author = {Jay, Aldan and Berlanga, Rafael},
  journal = {Algorithms},
  year = {2026},
  volume = {19},
  number = {2},
  eid = {101},
  doi = {10.3390/a19020101},
  url = {https://doi.org/10.3390/a19020101},
  note = {CC BY 4.0}
}

Summary and citation