• Tremendous growth in the influx of news related to traded assets in international financial markets via not only just print media but through real-time online sources such as social media. These news items are swiftly transformed into investors sentiment which in turn drives prices.

• The continuous release of financial news helps to update general investor’s information sets in relation to financial markets and influences general investor sentiment.

• Various commercial agencies have started developing their own financial news sentiment data sets which are used by investors and traders to support their algorithmic trading strategies.

• Some of them are

  • 1 Refinitiv Machine-ready news.
  • 2 MarketPsych Analytics
  • 3 Alexandria Technology
  • 4 Ravenpack Data

• This project is investigatiing 2 and 3.

• There is a growing body of research that argues that news items from different sources influence investor sentiment, and hence asset prices, asset price volatility and risk

• Tetlock (2007), Tetlock, SAAR-TSECHANSKY, & Macskassy (2008), Da, Engelberg, & Gao (2011); Barber & Odean (2008); diBartolomeo & Warrick (2005); Mitra & Mitra (2011).

• A recent development within the realm of these studies is a collation of research that explores indices that provide a proxy for human emotions such as fear, joy, and optimism. The studies examine whether emotional indices can explain innovations in financial asset prices which do not correspond to the expected behaviour of rational economic agents.

  • For instance, Griffith, Najand, & Shen (2020) investigate (utilizing the Thomson Reuters MarketPsych Indices – TRMI) whether proxies of fear, gloom, joy, and stress can be used to predict returns and volatility in the US equity market.
  • Papakyriakou, Sakkas, & Taoushianis (2019) explore whether unexpected announcements of company defaults in the US can induce asset price reactions in stock markets based in other geographic jurisdictions.

• This is a continuation of previous research where we looked at the RMN (Refinitiv Machine Ready News) data

  • See David E. Allen, McAleer, & Singh (2019); David E. Allen, McAleer, Powell, & Singh (2017) and D. Allen, McAleer, & Singh (2016)

• More recently, the mentioned datasets have been used in research utlising machine learning and econometric methods

• For example

  • Borovkova & Dijkstra (2018) incorporates machine learning techniques to undertake intraday forecasting of the European sharemarket using the TRMI as the main predictor. The study uses a Kalman filter to remove the noise of the sentiment signals (we can also use a moving average - just as good) and then employs Multi-Layer Perceptron (MLP) techniques to select the features of the proposed model. The study ends up using the Long Short-Term Memory (LSTM) network model to undertake the forecasts.
  • Audrino & Tetereva (2019) assesses returns for nominal exchange rates before, during, and after announcements of sovereign credit rating downgrades. The authors propose that in countries with weaker regulatory controls where informational leakages are more common, that currency depreciations occur before the announcements.
  • Obaid & Pukthuanthong (2022) applied machine learning to classify photos based on sentiment and TRMI dataset to examine news text to predict stock returns

This is not an exhaustive list (work in progress)

• Is there a significant relationship between Emotion based sentiment and stock price movement?
• Are topic based (FACTSET/Alexandria data) sentiment scores able to model asset prices?
• Was there a significant relationship between sentiment (from both TRMI and Alexandria) and stock/commodity/cryptocurrency price movements during COVID-19?
• There are possible questions about link between sentiment and volatility

Let’s look at these datasets briefly

• The data is in high frequency and has over 250 million rows (since 2000)
• There are stocks from over 105 countries but USA is the most dominant one
• The sentiment are generated from news from either Dow Jones news publication or other press release

• The news stories are categorised by topics

Let’s look at a one year sample (2021) from the DJIA of the USA

• The data isnt continuous

Sentiment scores for 1 day

• 1 Stock data

Sentiment Scores for 1 Stock (AAPL)

• The dataset provides various emotional indicators, such as, Anger, Fear, Joy, Gloom, Stress, Surprise etc., along with the overall positive and negative sentiments.
• The dataset also has indicators for price movement, such as Price Direction, Volatility etc.
• These indicators are available for 1 min, hourly and daily frequencies
• There are various asset classes available, including, stocks, currencies, cryptocurrencies.
• The sentiments are generated using news from news sites as well as social media.
• Here we look at Energy and Material commodity data for the year 2021 with 1 min frequency
• The dataset has around 4million rows and 43 fields

Data Source Type

Source Type

Asset types

Source Type

Sentiment Indicators

• Evaluation of various ML, Econometric methods to examine the research questions
• Experiment with weighting schemes to find a daily indicator
• Analysis around COVID-19, particularly in Commodities and Crypto.