Telegram News-Powered Investment Advisor: Enhancing Predictive Models for Stock Trading in German Market
This project was carried out as part of the TechLabs “Digital Shaper Program” in Düsseldorf (Summer Term 2023).
We predict stock price changes using scraped data from Telegram Channels in German
Abstract
Almost everybody who has stable income and savings wants to invest in something (e.g. shares). However not everyone knows the tactics on how to invest and at which moment to sell/buy shares. In order to help this group of people to invest we created an ML model which uses Telegram news in german language to add predictive power to existing ML models which are built just on time series data.
Introduction
The proposed project addresses these challenges by leveraging machine learning and natural language processing techniques to analyse Telegram news in the German language and integrate it with historical financial data. By doing so, it aims to provide investors with actionable insights, predictive analytics, and a user-friendly platform to make informed decisions, optimise their investment strategies, and navigate the complexities of the German stock market with confidence.
The proposed decision may help the beginner investors to solve such problems as:
— Information overload — there is too much information which also may be fake
— Lack of expertise
— Market volatility — the investor may get to know about some market downfalls as early as possible
— Language barrier — around 5% of Germany residents can’t speak German
— Risk Management — more accurate forecasts reduces risks of failure
Methods
To complete the project, we proceeded through these outlined stages:
1. We Chose the datasource for both financial and news data. We used Google Finance as a source for daily stock prices and German Wallpapers from Telegram.
2. We chose the following companies for our analysis since they had the most mentions in german news:
3. We connected both news and prices datasets into one dataset with prices, news text for each date and company
4. We converted news data into sentiments using the Bert model for German texts.
5. We run a basic predictive model for time series Price prediction (RidgeReg) and compared the results with the baseline (target model — with news sentiments, baseline — the one without sentiments)
Results
As can be seen on the following graph, the predictive power is stronger for companies which have a high amount of mentions in social media. However, for those companies which have just a few mentions in social media, it can be not really useful to include news into analysis. This results shows that if there was more news data stored for the analysis, the predictive power might be improved way better. Even with a small amount of news available it was possible to obtain improved metrics of quality (f1-score) compared to the baseline. In particular, following just a few Telegram Channels, people may make optimal decisions on how to invest in Lufthansa and Siemens shares.
GitHub repository: https://github.com/dariagerasimenko/stock_changes_german_news
