Paper: arXiv 2503.18096

Authors: Filip Stefaniuk, Robert Ślepaczuk

Abstract

The article investigates the usage of Informer architecture for building automated trading strategies for high frequency Bitcoin data. Three strategies using Informer model with different loss functions: Root Mean Squared Error (RMSE), Generalized Mean Absolute Directional Loss (GMADL) and Quantile loss, are proposed and evaluated against the Buy and Hold benchmark and two benchmark strategies based on technical indicators. The evaluation is conducted using data of various frequencies: 5 minute, 15 minute, and 30 minute intervals, over the 6 different periods. Although the Informer-based model with Quantile loss did not outperform the benchmark, two other models achieved better results. The performance of the model using RMSE loss worsens when used with higher frequency data while the model that uses novel GMADL loss function is benefiting from higher frequency data and when trained on 5 minute interval it beat all the other strategies on most of the testing periods. The primary contribution of this study is the application and assessment of the RMSE, GMADL, and Quantile loss functions with the Informer model to forecast future returns, subsequently using these forecasts to develop automated trading strategies. The research provides evidence that employing an Informer model trained with the GMADL loss function can result in superior trading outcomes compared to the buy-and-hold approach.

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 employs advanced machine learning architectures (Informer, Transformer-based) and novel loss functions (GMADL), representing moderate-to-high mathematical density. The empirical work is rigorous, utilizing high-frequency Bitcoin data across multiple timeframes and periods, with thorough backtesting against benchmarks, sensitivity analysis, and open-source implementation for reproducibility.

Research Flowchart

  flowchart TD
  A["Research Goal: Evaluate Informer architecture<br>for HFT Bitcoin trading strategies"] --> B["Data Processing"]
  B --> C["Model Training & Forecasting"]
  C --> D["Strategy Formulation"]
  D --> E["Performance Evaluation"]

  subgraph B ["Data & Inputs"]
      B1["Bitcoin OHLCV Data"]
      B2["High Frequency Intervals<br>5m, 15m, 30m"]
      B3["Six Distinct Time Periods"]
  end

  subgraph C ["Computational Process"]
      direction LR
      C1["Informer Architecture"] --> C2{"Loss Functions"}
      C2 --> C3["RMSE Loss"]
      C2 --> C4["GMADL Loss"]
      C2 --> C5["Quantile Loss"]
  end

  subgraph D ["Trading Logic"]
      D1["Generate Return Forecasts"]
      D2["Execute Buy/Sell Signals"]
      D3["Benchmarks: Buy & Hold<br>Technical Indicators"]
  end

  subgraph E ["Key Outcomes"]
      E1["RMSE: Performance degrades<br>with higher frequency data"]
      E2["GMADL: Benefits from high frequency<br>Best overall performance"]
      E3["Quantile: Underperformed<br>benchmarks"]
  end

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 three Informer-based trading strategies using different loss functions (RMSE, GMADL, Quantile) for Bitcoin.

Source sentence
Three strategies using Informer model with different loss functions: Root Mean Squared Error (RMSE), Generalized Mean Absolute Directional Loss (GMADL) and Quantile loss, are proposed

Assumes 4

  • Information Ratio is calculated assuming a risk‑free rate of zero. statistical
    Source sentence
    Information Ratio (IR*) The Sharpie ratio (Sharpe 1998) that is calculated under the assumption of a risk-free rate of zero.
  • There are 365 days in a year for converting intervals to annual figures. data
    Source sentence
    It is assumed that there are 365 days in a year, so for the datasets used in the research, this value is equal to Y5m = 105120, Y15m = 35040 and Y30m = 17520.
  • A flat transaction fee of 0.1% per trade is applied. execution
    Source sentence
    With transaction fee equal to 0.1% even if the asset price stays the same, already with log0.999(0.5) ≈692 incorrect position changes the portfolio value is halved
  • The standard spot fee rate on Binance is used as the transaction cost. execution
    Source sentence
    The value is equal to standard spot fee rate on Binance https://www.binance.com/en/fee/trading

Data

Backtest sample: BTC/USDT · crypto · intraday

  • BTC/USDT price data · Binance market-prices public
    Bitcoin
    Source sentence
    The strategies were evaluated independently over six periods on BTC/USDT data of various frequencies: 5 minutes, 15 minutes, and 30 minutes.

Recognised sources: Binance

Costs

Modeled Costs are part of the reported headline results.

0.1% per trade (standard Binance spot fee).

Source sentence
... The value is equal to standard spot fee rate on Binance

Finds 4

Reported performance (GMADL Informer (5‑minute)): annual return 115.9% · max drawdown 32.7%

  • The GMADL Informer strategy trained on 5‑minute data achieved the highest performance, beating all other strategies and the buy‑and‑hold benchmark. performance
    Source sentence
    The best performing strategy was the one based on predictions of the Informer model trained on 5-minute data with the GMADL loss function.
  • Eight strategies outperformed the buy‑and‑hold benchmark over the combined testing periods. performance
    Source sentence
    Hold benchmark, when compared throughout the testing period, i.e., combined results from all data windows.
  • Higher‑frequency data improved the GMADL Informer strategy but deteriorated the RMSE Informer strategy. performance
    Source sentence
    usage of high frequency data improved the performance of the GMADL Informer strategy, while deteriorated the performance of the strategy based on the Informer with RMSE loss function.
  • The GMADL Informer strategy achieved a VAL of 9.747, ARC of 115.88%, ASD of 54.44%, IR* of 2.129, and MD of 32.66% on 5‑minute data. performance
    Source sentence
    GMADL Informer 9.747 115.88% 54.44% 2.129 32.66% 7.552 846 44.80%

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 27 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (1 by independent authors)

Code

https://gitlab.com/FilipStefaniuk/wne-msc-thesis — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (77)

  1. 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 — → 2412.18405
  2. The efficiency of various types of input layers of LSTM model in investment strategies on S&P500 index (2022) — link
  3. Bitcoin: A Peer-to-Peer Electronic Cash System (2009)
  4. A Random Walk Down Wall Street (1973)
  5. Efficient Capital Markets: A Review of Theory and Empirical Work (1970)
  6. Efficient Capital Markets: II (1991) — link
  7. Reflections on the Efficient Market Hypothesis: 30 Years Later (2005) — link
  8. The Evolution of Technical Analysis: Financial Prediction from Babylonian Tablets to Bloomberg Terminals (2011) — link
and 69 more
  1. What Do We Know About The Profitability Of Technical Analysis? (2007) — link
  2. The Evolution Of Stock Market Efficiency Over Time: A Survey Of The Empirical Literature (2011) — link
  3. Stock return predictability and the adaptive markets hypothesis: Evidence from century-long U.S. data (2011) — link
  4. A Little Learning…: Confidence in Multicue Judgment Tasks (1978)
  5. Boys will be Boys: Gender, Overconfidence, and Common Stock Investment (2001) — link
  6. Learning To Be Overconfident (2001) — link
  7. Does the Stock Market Overreact? (1985)
  8. Prospect Theory: An Analysis of Decision under Risk (1979) — link
  9. Are Investors Reluctant to Realize Their Losses? (1996) — link
  10. Contagious Speculation and a Cure for Cancer: A Nonevent that Made Stock Prices Soar (2001) — link
  11. The Framing of Decisions and the Psychology of Choice (1981) — link
  12. Time Series Analysis: Forecasting and Control (1976)
  13. Stock price prediction using the ARIMA model (2014) — link
  14. A hybrid ARIMA and support vector machines model in stock price forecasting (2005) — link
  15. Time series forecasting using a hybrid ARIMA and neural network model (2003) — link
  16. Bitcoin Price Prediction: An ARIMA Approach (2018) — link
  17. Stock index forecasting based on a hybrid model (2012) — link
  18. Exponential smoothing model selection for forecasting (2006) — link
  19. A novel hybridization of artificial neural networks and ARIMA models for time series forecasting (2011) — link
  20. Long Short-Term Memory (1997) — link
  21. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation (2014) — link
  22. Deep learning (2015) — link
  23. Forecasting Economics and Financial Time Series: ARIMA vs. LSTM. (2018) — link
  24. Application of machine learning in algorithmic investment strategies on global stock markets (2023) — link
  25. Artificial neural networks approach to the forecast of stock market price movements (2016)
  26. Recurrent Neural Networks Approach to the Financial Forecast of Google Assets (2017) — link
  27. Predicting Prices Of S&P 500 Index Using Classical Methods and Recurrent Neural Networks (2024) — link
  28. Hybrid Deep Learning Model for Stock Price Prediction (2018) — link
  29. Deep learning with long short-term memory networks for financial market predictions (2017) — link
  30. Stock price prediction using LSTM, RNN and CNN-sliding window model (2017) — link
  31. A LSTM-based method for stock returns prediction: A case study of China stock market (2015) — link
  32. Stock market’s price movement prediction with LSTM neural networks (2017) — link
  33. Gradient amplification: An efficient way to train deep neural networks (2020) — link
  34. Neural Machine Translation by Jointly Learning to Align and Translate (2014) — link
  35. Effective Approaches to Attention-based Neural Machine Translation (2015) — link
  36. Attention Is All You Need (2017) — link
  37. GPT-4 Technical Report (2024) — link
  38. Gemini: A Family of Highly Capable Multimodal Models (2024) — link
  39. Deep Learning for Time Series Forecasting : Tutorial and Literature Survey (2023) — link
  40. DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks (2019) — link
  41. Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting (2020) — link
  42. MQTransformer: Multi-Horizon Forecasts with Context Dependent and Feedback-Aware Attention (2020) — link
  43. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021) — link
  44. Exploring the Advantages of Transformers for High-Frequency Trading — → 2302.13850
  45. Attention! Transformer with Sentiment on Cryptocurrencies Price Prediction (2022) — link
  46. Stock Price Prediction Based on Temporal Fusion Transformer (2021) — link
  47. Beyond Sentiment in Stock Price Prediction: Integrating News Sentiment and Investor Attention with Temporal Fusion Transformer (2024) — link
  48. Cryptocurrency Price Prediction with LSTM and Transformer Models Leveraging Momentum and Volatility Technical Indicators (2023) — link
  49. A Stock Price Prediction Method Based on BiLSTM and Improved Transformer (2023) — link
  50. An Informer Based Method for Stock Intraday Price Prediction (2023) — link
  51. Advanced Stock Price Prediction Using LSTM and Informer Models (2024) — link
  52. Stock and market index prediction using Informer network — → 2305.14382
  53. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018) — link
  54. Daily and intraday application of various architectures of the LSTM model in algorithmic investment strategies on Bitcoin and the S&P 500 Index (2022) — link
  55. Momentum and contrarian effects on the cryptocurrency market (2019) — link
  56. Technical analysis: power tools for active investors (2005)
  57. New Concepts in Technical Trading Systems (1978) — link
  58. Xception: Deep Learning with Depthwise Separable Convolutions (2017) — link
  59. Gradient Flow in Recurrent Nets: The Difficulty of Learning LongTerm Dependencies (2001) — link
  60. Layer Normalization (2016) — link
  61. 1D Convolutional Neural Networks and Applications: A Survey (2019) — link
  62. Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs) (2016) — link
  63. Max-pooling convolutional neural networks for vision-based hand gesture recognition (2011) — link
  64. Deep Learning using Rectified Linear Units (ReLU) (2019) — link
  65. Mean Absolute Directional Loss as a new loss function for machine learning problems in algorithmic investment strategies — → 2309.10546
  66. How algorithmic trading undermines efficiency in capital markets (2015)
  67. Markov processes over denumerable products of spaces, describing large systems of automata (1969)
  68. Bollinger on Bollinger Bands (2002) — link
  69. Applying Hybrid ARIMA-SGARCH in Algorithmic Investment Strategies on S&P500 Index (2022) — link

4 of 77 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.