Source-linked AI summary
Decision support from financial disclosures with deep neural networks and transfer learning
Mathias Kraus, Stefan Feuerriegel
TL;DR
Financial disclosures support investment and automated-trading decisions, but computerized systems must handle their complex content. This paper examines deep-learning-based decision support, including recurrent models and transfer learning, and reports higher directional accuracy than traditional machine learning.
Problem
Financial disclosures inform investment and automated-trading decisions, while computerized systems must process their content and context.
Method
The paper studies recurrent neural networks for financial decision support and further pre-trains word embeddings using transfer learning.
Results
Long short-term memory models outperform traditional bag-of-words machine-learning models, especially with transfer learning.
Takeaways & Limitations
Deep learning can enhance financial decision support by incorporating word order and context-related information.
Takeaways & Limitations
The approach requires a sufficiently large labeled dataset.
Abstract
from arXiv · showhide
Company disclosures greatly aid in the process of financial decision-making; therefore, they are consulted by financial investors and automated traders before exercising ownership in stocks. While humans are usually able to correctly interpret the content, the same is rarely true of computerized decision support systems, which struggle with the complexity and ambiguity of natural language. A possible remedy is represented by deep learning, which overcomes several shortcomings of traditional methods of text mining. For instance, recurrent neural networks, such as long short-term memories, employ hierarchical structures, together with a large number of hidden layers, to automatically extract features from ordered sequences of words and capture highly non-linear relationships such as context-dependent meanings. However, deep learning has only recently started to receive traction, possibly because its performance is largely untested. Hence, this paper studies the use of deep neural networks for financial decision support. We additionally experiment with transfer learning, in which we pre-train the network on a different corpus with a length of 139.1 million words. Our results reveal a higher directional accuracy as compared to traditional machine learning when predicting stock price movements in response to financial disclosures. Our work thereby helps to highlight the business value of deep learning and provides recommendations to practitioners and executives.
1. Introduction
Financial disclosures feed news-based decision support systems, but traditional text representations lose context and often yield unsatisfactory stock-movement predictions. This paper investigates deep sequence modeling and transfer learning to improve the prediction engine.
- Decision support systems rate disclosure content to identify stocks whose prices may rise or fall, informing subsequent trading and portfolio decisions.The prediction engine may return an expected excess return or a positive/negative price-direction label.
- Traditional bag-of-words models count word or n-gram frequencies while ignoring word order and context-dependent meaning.This representation maps variable-length documents to fixed-length input vectors.
- Bag-of-words limitations are associated with often unsatisfactory accuracy when forecasting stock-price movements from financial narratives.Negation examples illustrate why frequency counts can miss meaning changes caused by context.
- Deep sequence models process words in order, preserve contextual information, and model complex nonlinear relationships, potentially improving predictive power.The paper discusses recurrent neural networks and long short-term memory models as sequence-modeling variants.
- The paper asks whether deep sequence modeling improves short-term post-disclosure price prediction over bag-of-words methods and whether transfer learning further improves performance.It pre-trains word embeddings on a related financial-language corpus before transferring them to the study dataset.
- The study evaluates neural networks against established bag-of-words baselines and develops managerial recommendations for deep learning in financial decision support.The networks contain more than 500,000 parameters, and the paper presents findings for research and management.
2. Related work
Prior financial-news decision support commonly used traditional text representations and predictive models, while deep learning remained relatively underused. Earlier deep-learning applications faced limitations in context modeling or relied on heavily filtered data, motivating the use of recurrent neural networks and LSTMs.
- Financial-news decision support includes explanatory and predictive approaches, with predictive systems trained on historical news and stock prices to forecast unseen news outcomes.
- Directional accuracy is usually only marginally better than 50%, often only on a subset of the original dataset, underscoring the task’s difficulty.
- Traditional text-mining pipelines commonly convert financial text into bag-of-words representations, including tf-idf features, before training predictive models.
- The study reports superior performance for tf-idf bag-of-words models compared with word frequencies.
- Recurrent neural networks process raw text sequentially and learn context-sensitive features, but can suffer from vanishing gradients and short context dependencies.
- LSTMs address long sequential dependencies using forget gates and are therefore applied as the study’s deep-learning architecture for financial decision support.
- Earlier financial applications used recursive autoencoders or word tuples, but recursive autoencoders rarely learn long context dependencies and may require extensive dataset filtering.
- The authors report no known prior work using recurrent neural networks or LSTMs to improve decision support based on financial news.
3. Methods and materials
The study compares bag-of-words machine-learning baselines with deep learning methods for predicting stock-price movements after financial disclosures. It evaluates classification and regression using nominal and abnormal returns, a chronological disclosure dataset split, and recurrent architectures including RNNs and LSTMs.
- Research design: The study compares naive bag-of-words machine learning with deep learning techniques for predicting stock-price movements from financial disclosures.The framework evaluates deep learning architectures and transfer learning against traditional approaches.
- Research design: The experiments classify disclosures as predicting upward or downward movements and regress the magnitude of nominal and abnormal returns.Abnormal returns adjust for confounding market movements to isolate disclosure-related effects.
- Dataset: The corpus contains 13,135 German ad hoc announcements, with 10,895 observations retained after selecting disclosures published on trading days.The sample covers regulated English-language disclosures and uses daily stock-market data for abnormal-return calculations.
- Baseline models: Text baselines use preprocessing and tf-idf features, while sensitivity analysis evaluates adjacent-word n-grams and several linear, nonlinear, and kernel-based predictors.Preprocessing includes tokenization, lowercasing, punctuation and number removal, optional stemming, and numerical vectorization.
- Deep learning architectures: RNNs process words sequentially by updating hidden states from the previous state and current input, while LSTMs use gated cell states to retain information longer.The RNN input can use one-hot word or stem vectors, and the study also experiments with word embeddings for LSTMs.
4. Results
Deep learning outperforms traditional bag-of-words approaches for predicting both the direction and magnitude of stock-price movements, with transfer learning providing additional gains.
- Deep learning is superior to traditional bag-of-words approaches in predicting the direction and magnitude of stock price movements.
- 6.8 percentage points improvement over the naïve baseline is achieved by the LSTM with word embeddings.Word embeddings contribute 0.1 percentage points to balanced accuracy and elevate AUC by 0.5 percentage points.
- 7.1 percentage points total improvement is achieved by the LSTM with word embeddings after transfer learning.Transfer learning adds 0.8 percentage points in balanced accuracy, and this model performs best among all approaches.
- 0.021 AUC improvement over the strongest traditional model reaches statistical significance at the 0.05 level with transfer learning.The strongest traditional model has an AUC of 0.556.
- 5.34 % lower mean squared error than the naïve approach is achieved by the LSTM with pre-training and word embeddings.The reduction is 2.053 in mean squared error.
5. Discussion
The discussion connects the results to financial decision support, transfer learning, economic value, and comparability constraints across prior studies.
- Comparisons across studies are often not comparable because filtering, accuracy metrics, data splits, and time-series cross-validation differ.
- Deep learning outperforms the replicated SVM-based approach under the paper’s dataset, training processes, and evaluation metrics.
- Transfer learning first learns representations from a related dataset, then adapts the output layer to the target problem.
- Accurate financial-news prediction remains difficult because natural language is complex and historic prices explain only marginally future returns.
- 51 % accuracy increases the portfolio’s log-return from the random-guess scenario to 10 % over one year.The illustrative portfolio starts at $1,000, assumes 200 trading days, and includes one disclosure triggering a 5 % log-return.
- Practical adoption can be challenging because rapid software-library advances require considerable implementation adjustments.
- Large pre-assembled datasets remain difficult to obtain for natural-language transfer learning, motivating publication of trained networks and future dataset development.
6. Conclusion and outlook
The paper finds that deep neural networks can enhance financial decision support by modeling word order, context, and nonlinear relationships in financial disclosures. It also identifies transfer learning as beneficial, while noting that configuring these networks requires extensive parameter tuning and that further research should examine additional prediction horizons.
- Deep learning enhances financial decision support by explicitly incorporating word order and context-related information into predictive models.
- Long short-term memory models can outperform traditional machine-learning models based on bag-of-words approaches, especially when combined with transfer learning.
- Superior predictive performance depends on inferring context-dependent information from ordered word sequences and capturing highly nonlinear relationships.
- Deep neural network configuration remains challenging because favorable results require extensive parameter tuning.
- Future research should further examine deep-learning gains for news-based predictions, including intraday latency effects and longer-term outcomes.
- The authors expect deep learning to expand beyond academic research and the limited number of firms specializing in predictive analytics.