Papers with respect to NLP-based Financial Forecasting
June 4, 2021 · View on GitHub
-
Financial Indicators Prediction:
-
Beyond Accuracy:
Stock Price Movement Predicion
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Ding et al.(2014) | Using Structured Events to Predict Stock Price Movement: An Empirical Investigation | (1) S&P 500 from Yahoo!Finance, (2) Reuters and Bloomberg Their Open-sourced dataset (404) | Linear model, deep NN | Predicting S&P500 index: Acc: 59% MCC: 0.1683 Individual stock (Walmart) prediction: Acc: 70% MCC: 0.4679 | -/10/2006- -/11/2013 | Extract events from news based on OpenIE, Figure out the relation between financial events and stock market | EMNLP-14 |
| Ding et al.(2015) | Deep Learning for Event-Driven Stock Prediction | (1) S&P 500 from Yahoo!Finance,(2)Reuters and Bloomberg (404) | CNN | S&P500 index: Acc: 64.21%, MCC: 0.40. Individual stock: Acc: 65.48%, MCC: 0.41 | 01/10/2006-21/11/2013 | Events are extracted from news text.Predict uptrend and downtrend probability based on events. | IJCAI-2015 |
| Yang et al.(2018) | Explainable Text-Driven Neural Network for Stock Prediction | Same as Ding et al. 2015 | BiGRU+Multi-level Attention | Accuracy, MCC | -/10/2006- -/11/2013 | Accuracy: 0.62-0.74 varing from three companies | CCIS-18 (Best Paper) |
| Du and Ishii(2020) | Stock Embeddings Acquired from News Articles and Price History, and an Application to Portfolio Optimization | WSJ/Reuters+Bloomberg News | BiGRU | Accuracy, MCC | 2000-2015 | 2.80x/1.37x annual portfolio gain | ACL-20 |
| Duan et al.(2018)/(2019) | Deep Learning for Event-Driven Stock Prediction | Code&Data (Reuters) | Sentence-level Bi-LSTM | Cumulative Abnormal Return (CAR) | 2006-2015 | Target-Specific Representations of Financial News Documents | CoLING-18/TASLP |
| Merello(2018) | Investigating Timing and Impact of News on the Stock Market | Financial News | Attention model + LSTM + Dense | Accuracy, MCC | -/08/2017- -/03/2018 | Accuracy 62.8%, MCC 0.18 | ICDM Workshops 2018 |
| Li et al.(2020) | Modeling the Stock Relation with Graph Network for Overnight Stock Movement Prediction | Stocks within TPX500 and TPX100 index from Reuters Financial News | LSTM, relational graph convolutional network (LSTM-RGCN) | Accuracy | 01/01/2013- 28/09/2018 | Since using the relational graph convolutional network, the proposed network is able to predict the movement of stocks that is not directly associated with news | IJCAI-2020 |
| Dang et al.(2020) | “The Squawk Bot”∗: Joint Learning of Time Series and Text Data Modalities for Automated Financial Information Filtering | Stock time series: Apple (AAPL) and Google (GOOG) collected from Yahoo!Finance (open-sourced) | LSTM | Precision, Recall | -/-/2006- -/-/2013 | Propose MSIN that is able to discover- ing relevant documents in association with a given time series | IJCAI-2020 |
| Ding et al.(2020) | Hierarchical Multi-Scale Gaussian Transformer for Stock Movement Prediction | Daily quote data of all 3243 stocks in NASDAQ, 15-min quote data of 500 CSI-500 component stocks (NOT open source) | Gaussian Transformer which is an efficient architecture of RNN/CNN | Accuracy | 01/07/2010- 01/07/2019, 01/12/2015- 01/12-2019 | Apply Gaussian Transformer model on stock movement prediction, in which attention mechanism can help to capture extremely long-term dependencies of finance time series. | IJCAI-2020 |
| Liu et al.(2020) | Multi-scale Two-way Deep Neural Network for Stock Trend Prediction | FI-2010, CSI-2016 (NOT open source) | Deep NN (XGB, Recurrent CNN | Accuracy, F1 score | 2010, 2016 | Proposed a novel multi-scale model which achieves state-of-the-art performance on the benchmark dataset. | IJCAI-2020 |
| Feng et al.(2019) | Enhancing Stock Movement Prediction with Adversarial Training | (1)88 high-trade-volume-stocks in NASDAQ and NYSE (Xu and Cohen, 2018) (2)50 stocks in U.S. market, Code | LSTM, attention model | On dataset 1: Acc: 57.2%, MCC: 0.1483. On dataset 2: Acc: 53.05%, MCC: 0.0523. | 01/01/2014-01/01/2016, 01/01/2007-01/01/2016 | Regard features extracted from historical price as stochastic variables and fix the overfitting problem on training dataset | IJCAI-2019 |
| Si et al.(2013) | Exploiting Topic based Twitter Sentiment for Stock Prediction | (1)Twitter REST API, (2)S&P100 from Yahoo Finance | Dirichlet Process Mixture (DPM) model, Vector auto-regression | Accuracy: 68% | 02/11/2012-07/01/2013 | Topic-based time series sentiment analysis. Use historical index data and the topic-based sentiment time series to predict stock movement. | ACL-2013 |
| Hu et al.(2019) | Listening to Chaotic Whispers: A Deep Learning Framework for News-oriented Stock Trend Prediction | 1)2527 Chinese stocks (2)1,271,442 economic news from http: //www.eastmoney.com/ and http: //finance.sina.com.cn/ | Attention model, NN | Acc: 48% (tri-label classification problem) Gained 60% annualized return | -/-/2014- -/-/2017 | They pointed out three principles for news-oriented stock trend prediction, including sequential context dependency, di- verse influence and effective and efficient learning, by imitating the learning process of human. | WSDM-19 |
| Xu and Cohen(2018) | Stock Movement Prediction from Tweets and Historical Prices | (1)88 high-trade-volume-stocks in NASDAQ and NYSE (2) Twitter (also in the link above) | Attention model, NN | Acc: 58.23% MCC: 0.08 | 01/01/2014-01/01/2016 | Deep generative approach for stock movement, introduce a temporal auxiliary | ACL-18 |
| Chen & Wei(2018) | Incorporating Corporation Relationship via Graph Convolutional Neural Networks for Stock Price Prediction | CSI300 company from a public API tushare | Graph convolutional network, LSTM | Acc: 57.98% | 29/04/1027- 31/12/2017 | Predict the price movement in day “t” based on historical price from day “t-7” to “t-1” | CIKM 2018 |
| Kim et al.(2019) | HATS: A Hierarchical Graph Attention Network for Stock Movement Prediction | (1) 431 companies of S&P500 from Yahoo Finance, (2) Corporate relational data from Wikidata. Dataset, Code | Graph network, attention model, LSTM | Individual stock prediction: (tri-label), Acc: 39%, F1:0.32, Average daily return: 0.096 | 08/02/2013- 17/06/2019 | leverage corporate relational data and investigate different types of relation (75).Select useful relations automatically by NN. Make prediction through historical price data and correlation information. | |
| Ye et al.(2020) | Multi-Graph Convolutional Network for Relationship-Driven Stock Movement Prediction | CSI300 and CSI500 company from a public API tushare | Graph convolutional network, Gated recurrent units | CSI300L Acc:0.5754, F1:0.6981, MCC: 0.2171. CSI500: Acc: 0.5885, F1: 0.7199, MCC: 0.2377 | 01/06/2015- 05/12/2019 | The idea of this paper is similar with Kim et al. (2019). Differences: (1) predict market index or individual stock price (2) the binary class (3) how to calculate the correlation | |
| Keynes (1937) | The General Theory of Employment | \ | \ | \ | \ | The literature of stock market prediction was initiated by this paper. | Eco |
| Fama (1965) | The behavior of stock-market prices. | \ | \ | \ | \ | Establish the Efficient Market Hypothesis (EMH) | Eco |
| Bondt and Thaler (1985) | Does the Stock Market Overreact? | CRSP monthly return data from NYSE | \ | Cumulative Average Residuals | -/01/1926- -/12/1982 | Empirically prove that the stock price movement is predictable | Eco |
| Jegadeesh (1990) | Evidence of Predictable Behavior of Security Returns | CRSP monthly return data | Regression model | Strong negative/positive serial correlation | -/-/1929- -/-/1982 | Empirically prove that the stock price movement is predictable | Eco |
| Culter et al. (1998) | What moves stock prices? | Annual return of S&P stock price series | Regression model | R2 | -/-/1871- -/-/1986 | one of the first papers to investigate the relationship between news coverage and stock prices, since which empirical text analysis technology has been widely used across numerous disciplines | Eco |
| Brown and Cliff (2004） | Investor sentiment and the near-term stock market | NYSE | Kalmam filter, Vector Autoregression | R2 | -/03/1965- -/12/1998 | used sentiment surveys from companies and signal extraction techniques to derive investor sentiment from market indicators | Eco |
| Ito et al. (2017) | Development of sentiment indicators using both unlabeled and labeled posts | Yahoo Finance bulletin boards | Logistic Regression | Accuracy, F1 score, Pearson Correlation | 18/11/2014- 30/06/2016 | Propose a method for extracting sentiment indicators using unlabeled posts. Develop a sentiment indicator. | 2017 IEEE Symposium Series on Computational Intelligence (SSCI) |
| Ito et al. (2017) | Development of an Interpretable Neural Network Model for Creation of Polarity Concept Dictionaries | 1. Financial news of companies from Tokyo Stock Exchange from Reuters 2. Yahoo Finance Board | Deep Neural Network | F1 score | -/01/2013- -/12/2015 | Develop an interpretable NN to extract the sentiment polarity score from documents. And proposed a concept dictionary. | 2017 ICDM Workshop |
| Ito et al. (2016) | Polarity propagation of financial terms for market trend analyses using news articles | 1. Financial news of companies from Tokyo Stock Exchange from Reuters 2. Yahoo Finance Board | Deep Neural Network, SVM | Precision, recall, f1 score | -/01/2014- -/12/2014 | proposed a new text-mining method for giving a polarity score to a new word | 2016 IEEE Congress on Evolutionary Computation (CEC) |
Stock Return Forecasting
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Yang et al. (2019) | Leveraging BERT to Improve the FEARS Index for Stock Forecasting | Data (Google Search Trend) | BERT+Online Training | MSE | Weekly Data from 2004 to 2015 | Combining the FEAR index (calculated by the google search volumn) with the semantic information (key words embedding) | FinNLP-2019 |
| Kelly (2018) | Estimating the impact of domain-specific news sentiment on financial assets | Dow Jones Industrial Average and West Texas Intermediate crude oil | Rolling window regression,Vector autoregression | Stock market: AR increased by 4.2%, MD decreased by 4.4%, Crude oil market: AR increased by 25.6%, MD decreased by 28.8% | 18/02/1998-31/07/2015 | present a method and implementation that analyses the content of news using multiple dictionaries that accounts for the specific use of terminology in a given domain | Knowledge-Based Systems |
| Qin et al. (2017) | A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction | (1) SML 2010, (2)NASDAQ 100 stock | LSTM, attention model | On NASDAQ dataset, MAE 0.21, MAPE 0.43, RMSE 0.31 | 26/07/2016- 22/12/2016 | Stage1: attention-based feature extraction. Stage2: temporal attention mechanism to select relevant hidden states. | IJCAI-2017 |
| Huang et al. (2017) | Forecasting stock returns in good and bad times: The role of market states | \ | \ | \ | \ | The goal in this paper is to show that stock returns are predictable in good times, rather than how to achieve the maximum predictability | Eco |
| Yuan (2015) | Market-wide attention, trading, and stock returns | Trim Tabs Financial | \ | \ | -/02/1998- -/12-2005 | analyze the ability of record-breaking events for the Dow index and front-page articles about the stock market to predict trading patterns and market returns | Eco |
| Oliveira et al. (2013) | Some Experiments on Modeling Stock Market Behavior Using Investor Sentiment Analysis and Posting Volume from Twitter | (1)All daily tweets from Twitter,(2)Stock market daily variables from Reuters (denied to access) | Regression model | Returns: R2 0.2, Trading volume: R2 0.41, Volatility: R2 0.67 | 24/12/2012- 08/02/2013 | Confirm the return is unpredictable based on the sentiment indicator.But they think trading volume and volatility are relevant to the tweets volume. | WIMS 2013 |
| Schumaker and Chen (2009) | Textual Analysis of Stock Market Prediction Using Breaking Financial News: The AZFinText System | (1) S&P 500 (2) 9211 financial news articles from Yahoo Finance | SVM | Directional accuracy: 57%, MSE: 0.04216, Simulated trading: 2.06% return | 26/10/2005- 28/11/2005 | They show that the model containing both article terms and stock price at the time of article release had the best performance | ACM Transactions on Information Systems |
| Duan et al. (2018) | Learning Target-Specific Representations of Financial News Documents For Cumulative Abnormal Return Prediction | Reuters | Neural attention, Bi-directional LSTM | AUC | 09/2006-12/2015 | Target-specific document representation model which disregards noise. Uses full news article for stock prediction. | COLING-2018 |
Portfolio Management
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Malandri(2018) | Public Mood–Driven Asset Allocation: the Importance of Financial Sentiment in Portfolio Management | NYSE companies | Random forest classifier, Multi-layer Perceptron, LSTM | Trading simulation | 24/01/2012-02/06/2017 | Discuss how public mood affect portfolio management | Cognitive Computation |
| Frank Z. X. (2018) | Intelligent Asset Allocation via Market Sentiment Views | StockTwits | Recurrent Neural Network | Trading simulation of one period portfolio asset allocation (5 stocks) | 08/14/2017-16/11/2017 | A novel neural network design, built upon an ensemble of evolving clustering and long short-term memory, is used to formalize sentiment information into market views.These views are later integrated into modern portfolio theory through a Bayesian approach. | IEEE Computational Intelligence Magazine |
| Picasso（2019) | Technical analysis and sentiment embeddings for market trend prediction | NASDAQ 100 companies | RF, SVM, feed forward NN | Trading simulation (20 stocks) | 03/07/2017-14/06/2018 | Maximum 85.2% Annualized Return | Expert Systems with Applications |
| Zhu et al. (2020) | Online Portfolio Selection with Cardinality Constraint and Transaction Costs based on Contextual Bandit | New York Stock Exchange (NYSE), Toronto Stock Exchange (TSE), S&P500 and Dow Jones 30 composite stocks (DJIA). (open source) | Asset combination selection algorithm: Lazy Exp4. Allocation algorithm: Transaction Costs-aware Gradient Projection (TCGP) | Maximum Drawdown and Cumulative Return on all four datasets | 03/06/1962-31/12/1984, 04/01/1994- 31/12/1998, 14/01/2001- 14/01/2003, 11/02/2013- 07/02/2018 | Propose an online portfolio selection method taking the Cardinality Constrains and Transaction Costs into account | IJCAI-2020 |
| Nakagawa et al. (2020) | RM-CVaR: Regularized Multiple β-CVaR Portfolio | Fama and French (FF) dataset (select FF25 and FF48) (NOT open-sourced) | Regularized Multiple β-CVaR portfolio | Annualized return increased by 7%, Maximum Drawdown decreased by 15% | -/01/1989- -/12/2018 | Use Conditional Value-at-Risk (CVaR) as the risk measure (formulated by single β ) and raise a new method to tackle the problem that small change of β causes huge change of portfolio structure | IJCAI-2020 |
| Cai (2020) | Vector Autoregressive Weighting Reversion Strategy for Online Portfolio Selection | NYSE_O, NYSE_N, DJIA, TSE (NOT open-sourced) | Vector Autoregressive moving-average (VARMA) | Cumulative Wealth | 03/07/1962- 30/06/2010 | To improve the performance of reversion based online portfolio selection strategy | IJCAI-2020 |
| Xu et al. (2020) | Relation-Aware Transformer for Portfolio Policy Learning | Crypto-A with 12 assets, Crypto-B with 37 assets, S&P-500 with 506 assets | Sequential Model | Cumulative return | -/02/2013- -/11/2019 | Tackle the problem that current portfolio policies are not able to capture the 1) sequential patterns of assets price series, 2) price correlation among multiple assets | IJCAI-2020 |
| Huang & Li (2020) | A Two-level Reinforcement Learning Algorithm for Ambiguous Mean-variance Portfolio Selection Problem | experimental work is done in MATLAB by the Monte Carlo simulations of returns from a Gaussian mixture model (GMM) as the underlying mixture distribution | Mean-Variance portfolio policy, Progressive Hedging Algorithm (PHA) | The Two-Layer framework produces | \ | Assume the statistics of assets’ returns are unknown to the investors, propose a portfolio management framework | IJCAI-2020 |
| Pun et al. (2020) | Financial Thought Experiment: A GAN-based Approach to Vast Robust Portfolio Selection | An empirical dataset on S&P 500 index (NOT open-sourced) | Generative Adversarial Network (GAN), and a regression network | In bearish market, GANr has the lowest annualized risk. In bullish market, GANr’s performance is similar with two benchmarks. | 01/12/2006-30/06/2009, 31/12/2015-29/12/2017 | Build an adversarial network to mimic the trading behaviors in the real-world, and results in a robust portfolio. | IJCAI-2020 |
| Lee et al. (2020) | MAPS: Multi-agent Reinforcement Learning-based Portfolio Management System | 18 year’s data of 3000 US companies (list of companies from Russell 3000 index) (NOT open source) | MLP encoder, Q-network | The value of MAPS portfolio is twice as the other baseline models | -/-/2012- -/-/2018 | Each agent stands for an investor. Propose a model that guides the agents act as diversely as possible while maximum their own returns | IJCAI-2020 |
| Markowitz (1954) | Portfolio Selection (unaccessible now) | \ | \ | \ | \ | Introduce Mean-Variance Theory of portfolio management | Eco |
| Kelly (1956) | A New Interpretation of Information Rate. | \ | \ | \ | \ | Capital Growth Theory. | Eco |
| Cover (1991) | UNIVERSAL PORTFOLIOS | \ | \ | \ | \ | Propose universal portfolios algorithm. The universal portfolio algorithm is a portfolio selection algorithm from the field of machine learning and information theory. The algorithm learns adaptively from historical data and maximizes the log-optimal growth rate in the long run. | Eco |
| Markowitz et al.(2000) | Mean-variance analysis in portfolio choice and capital markets | \ | \ | \ | \ | Introduce Mean-Variance Theory of portfolio management | Eco |
Trading Strategy Analysis
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Zhong et al. (2020) | Data-Driven Market-Making via Model-Free Learning | Chicago Mercantile Exchange (CME)’s Globex electronic trading platform (NOT open source) | Stochastic iterative method (Q- learning) | Proposed Q-learning algorithm outperforms two benchmark algorithms and the firm’s trading strategy | 01/01/2019- 31/12/2019 | Study when a market-making firm should place orders to maximize their expected net profit, while also constraining risk | Eco |
| Lin & Beling (2020) | An End-to-End Optimal Trade Execution Framework based on Proximal Policy Optimization | One-year millisecond Trade and Quote (TAQ) data of 14 stocks (NOT open-sourced) | Fully Connected Network, LSTM | Mean of Implementation Shortfall (IS), standard deviation of IS, and Gain-Loss Ratio (GLR) | -/01/2018- -/12/2018 | Based on Limit Order Book (LOB) information such as bid or ask prices, make trading decision directly without manually attributes | IJCAI-2020 |
| Spooner & Savani (2020) | Robust Market Making via Adversarial Reinforcement Learning | \ | Adversarial Reinforcement Learning | The resulting performance shows an improvement in the Sharpe ratio of 0.27 and lower variance on terminal wealth | \ | Use Adversarial Reinforcement Learning to product market making agents | IJCAI-2020 |
| Poli et al. (2020) | WATTNet: Learning to Trade FX via Hierarchical Spatio-Temporal Representation of Highly Multivariate Time Series | Collect 6 major NDF markets: US Dollar – Chinese Yuan, US Dollar – Indonesian Rupiah, US Dollar – Indian Rupee, US Dollar – Philippine Peso, US Dollar – Taiwan Dollar (NOT open-sourced) | NN | 219.1 Return over Investment in USDCNY market, whereas LSTM model got 74.3 | 10/09/2013-17/06/2019 | Focus on non-deliverable forward (NDF) selection, which is a derivatives contract used in foreign ex- change (FX) trading | IJCAI-2020 |
Measuring Forecasting Skill
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Zong et al. (2020) | Measuring Forecasting Skill from Text | Geopolitical Forecasting Data, Earnings Call Data | BERT-base | \ | 2014-2018 | They presented the first study of connections between people’s forecasting skill and language used to justify their predictions. | ACL-2020 |
Interpretable Model
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Ito et al. (2020) | Contextual Sentiment Neural Network for Document Sentiment Analysis. | 1.EcoRevs I and II. 2. Yahoo review 3. Sentiment 140 | Deep NN, Contextual Sentiment Neural Network (CSNN) | Macro F1 score | \ | To improve the interpretability of NN, they propose a novel initialization propagation (IP) learning to replace BP algorithm. | Data Science and Engineering |
| Ito et al. (2020) | SSNN: Sentiment Shift Neural Network | 1. EcoRevs I and II. 2. Yahoo review 3. Sentiment 140 | Sentiment Shift Neural Network | Macro F1 score | \ | Proposed a Joint Sentiment Propagation (JSP) learning to realize the interpretability of neural network layers | Proceedings of the 2020 SIAM International Conference on Data Mining |
| Ito et al. (2020) | Word-Level Contextual Sentiment Analysis with Interpretability | 1. EcoRevs I and II. 2. Yahoo review 3. Sentiment 140 | Sentiment Interpretable Neural Network (SINN), | Macro F1 score | \ | Propose Lexical Initialization Learning to improve the interpretability of NN. | Proceedings of the AAAI Conference on Artificial Intelligence |
| Ito et al. (2019) | CSNN: Contextual Sentiment Neural Network | 1. EcoRevs I and II. 2. Yahoo review 3. Sentiment 140 | Contextual Sentiment Neural Network (CSNN) | Macro F1 score, Pearson Correlation Coefficient | \ | To improve the interpretability of NN, they propose a novel initialization propagation (IP) learning to replace BP algorithm. | 2019 IEEE International Conference on Data Mining (ICDM) |
| Nakagawa et al. (2019) | Deep Recurrent Factor Model: Interpretable Non-Linear and Time-Varying Multi-Factor Model | 1. TOPIX 500 index from Tokyo Stock Exchange, 2. Nikkei Portfolio Master (NPM) and Bloomberg | LSTM-LRP | MAE, RMSE, annualized return, volatility, Sharpe ratio | -/12/1990- -/03/2015 | By combining the Layer-wise Relevance Propagation (LRP) with LSTM, they improved the interpretability of model. | AAAI-19 Workshop on Network Interpretability for Deep Learning |
Discourse Analysis
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Ito et al. (2020) | Learning Company Embeddings from Annual Reports for Fine-grained Industry Characterization | 1.10K reports of companies in US stock market(in English) 2.10K reports of companies in Tokyo Stock Exchange(in Japanese) | BERT-base-uncased model (for English text), Japanese pre-trained model (for Japanese text) | Related Company Extraction Test, Theme-based Extraction Test | 2018, 2019 | propose to learn vector representations of companies based on their annual reports | Proceedings of the Second Workshop on Financial Technology and Natural Language Processing |
| Ito et al. (2019) (cannot download from UCD library) | Segment Information Extraction From Financial Annual Reports Using Neural Network | \ | \ | \ | \ | \ | 2019 Annual Conference of the Japanese Society for Artificial Intelligence |
Text Visualization
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Ito et al. (2018) | GINN: Gradient Interpretable Neural Networks for Visualizing Financial Texts | 1.Yahoo Finance Board, 2. Reuters financial news, 3. Stock codes from Tokyo Stock Exchange | Gradient Interpretable Neural Network (GINN) | Macro F1 score | -/01/2007- -/12/2016 | the GINN can visualize important concepts given in various sentence contexts. Such visualization helps nonexperts easily understand financial documents. | International Journal of Data Science and Analytics |
| Ito et al. (2019) (cannot download from UCD library) | Concept Cloud-based Sentiment Visualization for Financial Reviews | \ | \ | \ | \ | \ | 2019 The International Conference on Decision Economics |
| Ito et al. (2019) | Word-level Sentiment Visualizer for Financial Documents | 1.Current economy watchers survey, 2.Synthetic dataset, 3.Yahoo Dataset, 4.Manually created word polarity lists. | LRP-RNN, Bidirectional LSTM | Macro F1 score | \ | Proposed Layer-wise Relevance Propagation (LRP) for word-level sentiment. Proposed two frameworks LWSV and GWSV for financial text-visualization. | 2019 IEEE Conference on Computational Intelligence for Financial Engineering (CIFEr) |
| Ito et al. (2018) (cannot download from UCD library) | Text-visualizing Neural Network Model: Understanding Online Financial Textual Data | \ | \ | \ | \ | \ | Pacific-Asia Conference on Knowledge Discovery and Data Mining |
Other
| Reference | Paper | Data Source | Model | Evaluation Metric(s) | Time Period | Contributions | Venue |
|---|---|---|---|---|---|---|---|
| Zamani et al. (2017) | Using Twitter Language to Predict the Real Estate Market | Census Bureau, Zillow, Twitter | Residualised Control Regression | \ | 2011-2013 | Shows twitter data can be predictive of real estate. Residualised control approach to multi-modal features. | EACL-2017 |