Paper: arXiv 2404.01866
Authors: Bartosz Bieganowski, Robert Slepaczuk
Abstract
This paper investigates the enhancement of financial time series forecasting with the use of neural networks through supervised autoencoders, aiming to improve investment strategy performance. It specifically examines the impact of noise augmentation and triple barrier labeling on risk-adjusted returns, using the Sharpe and Information Ratios. The study focuses on the S&P 500 index, EUR/USD, and BTC/USD as the traded assets from January 1, 2010, to April 30, 2022. Findings indicate that supervised autoencoders, with balanced noise augmentation and bottleneck size, significantly boost strategy effectiveness. However, excessive noise and large bottleneck sizes can impair performance, highlighting the importance of precise parameter tuning. This paper also presents a derivation of a novel optimization metric that can be used with triple barrier labeling. The results of this study have substantial policy implications, suggesting that financial institutions and regulators could leverage techniques presented to enhance market stability and investor protection, while also encouraging more informed and strategic investment approaches in various financial sectors.
Complexity vs Empirical Score
- Math Complexity: 3.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper employs well-defined neural network architectures and hyperparameter tuning with a clear walk-forward backtesting protocol on specific assets, demonstrating strong empirical rigor. However, the mathematical complexity is moderate, primarily involving standard deep learning concepts and statistical performance metrics rather than dense theoretical derivations.
Research Flowchart
flowchart TD A["Research Goal:<br>Enhance Financial Time Series<br>Forecasting via Supervised Autoencoders"] --> B["Data & Inputs:<br>S&P 500, EUR/USD, BTC/USD<br>Jan 2010 - Apr 2022"] B --> C["Methodology:<br>Triple Barrier Labeling<br>& Noise Augmentation"] C --> D["Computational Process:<br>Supervised Autoencoder MLP<br>(Tuning Bottleneck/Noise)"] D --> E["Evaluation:<br>Sharpe & Information Ratios"] E --> F["Key Findings/Outcomes:<br>Optimal Noise/Bottleneck Boosts Returns<br>Excessive Noise Impairs Performance<br>Novel Optimization Metric Derivation"]
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 using supervised autoencoders with noise augmentation and triple barrier labeling to improve financial time series forecasting and strategy performance.
Source sentence
This paper investigates the enhancement of financial time series forecasting with the use of neural networks through supervised autoencoders, aiming to improve investment strategy performance. It specifically examines the impact of noise augmentation and triple barrier labeling on risk-adjusted returns
Assumes 6
- Zero risk‑free rate is assumed for the Information Ratio calculation. model
Source sentence
Because we assume a zero-rate risk-free rate, instead of Sharpe Ration we will define IR
- Stop‑losses and take‑profits from triple barrier labeling are executed immediately and perfectly. execution
Source sentence
The study also assumes that stop-losses and take profits from triple barrier labeling are executed immediately and perfectly
- Historical price data are representative for evaluating future strategy performance. market
Source sentence
the findings are based on historical data, and as such, their predictive power in relation to future performance should be considered with caution
- Features must be stationary for supervised learning to be reliable. data
Source sentence
Supervised learning requires stationary features because it involves mapping an unseen observation to labeled examples to predict the observation’s label
Show 2 more
- Transaction costs are fixed percentages per trade side with no market impact. execution
Source sentence
Transaction cost for S&P 500 Index and EUR/USD amounts to 0.005%, for Bitcoin it is 0.1%
- The model’s Gaussian noise augmentation is applied to features before training. model
Source sentence
Data augmentation has found a valuable place in the domain of time series analysis as well
Data
- S&P 500 index daily prices market-prices public
Source sentence
The study focuses on the S&P 500 index, EUR/USD, and BTC/USD as the traded assets from January 1, 2010, to April 30, 2022
- EUR/USD daily prices market-prices public
Source sentence
The study focuses on the S&P 500 index, EUR/USD, and BTC/USD as the traded assets from January 1, 2010, to April 30, 2022
- BTC/USD daily prices market-prices public
Source sentence
The study focuses on the S&P 500 index, EUR/USD, and BTC/USD as the traded assets from January 1, 2010, to April 30, 2022
- Initial Claims (ICSA) series macro public
Source sentence
the first one is the seasonally adjusted Initial Claims (ICSA), retrieved from the Federal Reserve Bank of St. Louis, FRED
Costs
Modeled
Transaction cost for S&P 500 Index and EUR/USD amounts to 0.005%, for Bitcoin it is 0.1%
Source sentence
Transaction cost for S&P 500 Index and EUR/USD amounts to 0.005%, for Bitcoin it is 0.1%
Finds 5
- Supervised autoencoders significantly improve strategy performance, with an optimal balance of Gaussian noise level and bottleneck size. performance
Source sentence
Through rigorous testing and sensitivity analysis, we demonstrated that employing supervised autoencoders significantly enhanced the performance metrics of the strategy
- Excessive noise and overly large bottleneck sizes degrade performance, likely due to overfitting and added noise. negative
Source sentence
an increase in the level of noise and bottleneck size beyond certain thresholds led to a decrease in performance
- Approach 3 (with noise augmentation and denoising autoencoder) outperformed Approaches 1 and 2 across all bar lengths in Information Ratio of an equally weighted portfolio. performance
Source sentence
Approach 3 outperformed Approaches 1 and 2 across all bar lengths in terms of the Information Ratio of an equally weighted portfolio
- Triple barrier labeling generally outperforms simple direction classification, especially on 15‑minute and 30‑minute bars. performance
Source sentence
Triple barrier labeling generally outperformed simple labeling due to its ability to handle market noise better
- In the 5‑minute bar setting, the equally‑weighted portfolio showed a marginal cumulative loss of -2.16% with an information ratio of -0.11. performance
Source sentence
The equally-weighted portfolio, however, resulted in a marginal cumulative loss of -2.16%, an information ratio of -0.11
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 20 related papers). Coverage is partial, so this is neither replication nor refutation.
3 papers in the archive : Is attention truly all we need? An empirical study of asset …
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 (42)
- Evaluating multiple classifiers for stock price direction prediction (2015) — link
- The heretics of finance: conversations with leading practitioners of technical analysis (2009) — link
- A Random Walk Down Wall Street (1973)
- Handbook of the Economics of Finance (2013)
- Artificial neural networks in business: Two decades of research (2015) — link
- Stock Price Prediction Using the ARIMA Model (2014) — link
- Advances in neural information processing systems 7 (1997) — link
- Exponential smoothing model selection for forecasting (2005) — link
and 34 more
- Artificial neural networks: a tutorial (1996) — link
- Introduction to financial forecasting (1996) — link
- AN INTRODUCTION TO LONG‐MEMORY TIME SERIES MODELS AND FRACTIONAL DIFFERENCING (1980) — link
- Long Short-Term Memory (1997) — link
- Predicting the Brazilian stock market through neural networks and adaptive exponential smoothing methods (2009) — link
- Comparison of support-vector machines and back propagation neural networks in forecasting the six major Asian stock markets (2006) — link
- Learning to Forget: Continual Prediction with LSTM (2000) — link
- Reflections on the Efficient Market Hypothesis: 30 Years Later (2005) — link
- Application of evolutionary computation for rule discovery in stock algorithmic trading: A literature review (2015) — link
- Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500 (2016) — link
- Deep learning with long short-term memory networks for financial market predictions (2017) — link
- Algorithmic financial trading with deep convolutional neural networks: Time series to image conversion approach (2018) — link
- Momentum and contrarian effects on the cryptocurrency market (2019) — link
- A Comparison of ARIMA and LSTM in Forecasting Time Series (2018) — link
- Hybrid Deep Learning Model for Stock Price Prediction (2018) — link
- A Comparative Study of LSTM and DNN for Stock Market Forecasting (2018) — link
- Bitcoin Price Prediction: An ARIMA Approach (2019) — link
- Robustness of Support Vector Machines in Algorithmic Trading on Cryptocurrency Market (2018) — link
- Advances in Financial Machine Learning (2018)
- A Comparative Study of Bitcoin Price Prediction Using Deep Learning (2019) — link
- Intelligent forecasting with machine learning trading systems in chaotic intraday Bitcoin market (2020) — link
- Deep learning for financial applications : A survey (2020) — link
- Generación de series de tiempo financieras sintéticas para “data augmentation” usando redes neuronales generativas adversarias (GAN) (2021)
- Profitability of technical stock trading: Has it moved from daily to intraday data? (2008) — link
- Financial Forecasting with Machine Learning: Price Vs Return (2021) — link
- LSTM in Algorithmic Investment Strategies on BTC and S&P500 Index (2022) — link
- NetVIX — A network volatility index of financial markets (2022) — link
- Time Series Analysis: Forecasting and Control (1977) — link
- Efficient Capital Markets: A Review of Theory and Empirical Work (1970) — link
- Comparing Predictive Accuracy (2002) — link
- Daily and intraday application of various architectures of the LSTM model in algorithmic investment strategies on Bitcoin and the S&P 500 Index (2023) — link
- Forecasting cryptocurrencies volatility using statistical and machine learning methods: A comparative study (2023) — link
- A Rank Graduation Box for SAFE AI (2024) — link
- Informer in Algorithmic Investment Strategies on High Frequency Bitcoin Data — → 2503.18096
1 of 42 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.