Using FinBert to predict / test for stock price movements based on news sentiment scores

Link to Dataset : https://www.kaggle.com/datasets/miguelaenlle/massive-stock-news-analysis-db-for-nlpbacktests

The idea is to use the dataset of stock news: headline, links, date, etc. to predict (or test) the prices movements of those stocks (increase or decrease or remain stable) based on sentiment score of the news using FinBert.

FinBert is an opensource pre trained Natural Language Processing (NLP) model, that has been specifically trained on Financial data, and outperforms almost all other NLP techniques for financial sentiment analysis.

Currently: Sentiment scored on news headlines not news content

---

RESULTS:

Result for df ( null values removed )

   precision    recall  f1-score   support

   decrease       0.09      0.23      0.13        31
   increase       0.06      0.19      0.09        26
     stable       0.86      0.60      0.71       355

    accuracy                           0.55       412
   macro avg       0.34      0.34      0.31       412
weighted avg       0.75      0.55      0.63       412

Using Headline Sentiment alone is quite good(?) to predict for stable days i.e. stock price remaining between +/- 0.20% (precision, recall and f1)

Note: supports are very high for 'stable' but then we only modeled/predict based on our threshold of sentiment scores

Metric	Desc.
Accuracy	true prositive + true negative / total
precision	true positives / actual results
recall	true positve / predicted resuts
f1-score	harmonic mean of precision and recall

---

Good Stability predictions:

• Usefull for predicting higher / lower volatility during the trading days (closing prices)
• Money can be made by using Option Strategies such as:
  1. Iron Condor for less volatility (within -0.20% +0.20% price change range)
  2. Straddle for high volatiolity (lower or higher than +/- 0.20% price change)

---

Result for filtered df ( only for stocks that made over -/+0.20% ):
Removing 'stable' days to just look at 'increase' and 'decrease' performance:

   precision    recall  f1-score   support

    decrease       0.58      0.23      0.33        31
    increase       0.56      0.19      0.29        26
      stable       0.00      0.00      0.00         0

    accuracy                           0.21        57
   macro avg       0.38      0.14      0.20        57
weighted avg       0.57      0.21      0.31        57

Not good enough.

---

Next Steps:

• Scrape each stock news content with the links in dataset
• Use the news body content for sentiments
• Compile results of predictions