FinBERT

Sentiment scores are computed by the weighted sum of the sentiment labels (positive, negative, or neutral) provided by the FinBERT model, and the labels' associated confidence levels. Sentiment score distributions for the following companies are shown as examples.

AMZN Sentiment

AMZN

CVS Sentiment

CVS

GOOG Sentiment

GOOG

Figure: Sentiment Score Distribution of Stocks

CDNOD Result

AT&T (T): Treasury Stock directly impacts AT&T's closing price, driven by Inventory, Prop- erty, Plant & Equipment, and Capital Lease Obligations. These factors shape free cash flow, enabling stock buybacks that reduce outstanding shares and boost the stock price. Since AT&T relies heavily on infrastructure to deliver services, efficient capital allocation and liq- uidity management are key to sustaining shareholder value.

Amazon (AMZN): Operating Income and M1 Money Stock directly impact Amazon's clos- ing price, driven by M2 Money Stock, Cash Equivalents, Investment Income, Capital Expen- diture, and Cash Flows. These factors shape free cash flow. Given Amazon's infrastructure- heavy model, efficient capital allocation and liquidity management are crucial for sustaining shareholder value.

Google (GOOG): M2 Money Stock, Non-Operating Income, Income Tax Expense, and CPI directly impact Google's closing price, driven by M1 Money Stock, PPI, stock repurchases, and the Unemployment Rate. These factors affect Google's ability to invest, control ex- penses, and execute buybacks. With its reliance on infrastructure and ad revenue, efficient liquidity management is essential for sustaining shareholder value.

CVS (CVS): The Unemployment Rate, Inventory, Total Non-Current Liabilities, and Cur- rent Debt directly impact CVS's stock price, driven by Capital Expenditure, Debt-Related Metrics, and Security Expenditure. With substantial debt, these factors reflect CVS’s ability to manage financial obligations, sustain operations, and invest in growth, ultimately influ- encing its stock performance. As a retail and healthcare company, CVS heavily relies on human resources, making labor market conditions a key factor in its operational efficiency and profitability.

Abbott Laboratories (ABT): Changes in Operating Cash Flow, Interest, and Debt Expense directly impact ABT's stock price, originating from Changes in Receivables, Current Debt, Operating Cash Flow, and Depreciation, Depletion, and Amortization (DDA). As a health- care company reliant on drug innovation, these factors influence cash availability, borrow- ing costs, and profitability, all of which are critical to Abbott's ability to fund Research and Development, expand operations, and return capital to shareholders.

Amgen (AMGN): Changes in Operating Assets and Month Labels directly impact Amgen's stock price, as fluctuations in inventory, receivables, and seasonal trends in drug sales can affect revenue recognition and investor sentiment

AMZN cdnod

AMZN

GOOG cdnod

GOOG

T Sentiment

T

AMGN cdnod

AMGN

CVS cdnod

CVS

ABT cdnod

ABT

Figure: Sentiment Score Distribution of Stocks

Model Performance Comparison

Overall Performance Across Companies

Our evaluation shows that both models have their strengths depending on the company and market conditions:

  • DeepAR excels with: ABT, AT&T, AMGN, and CVS (better error metrics)
  • Fusion Layer performs better with: AMZN and GOOG (lower prediction errors)
  • Direction accuracy: Fusion Layer achieved 100% for ABT and AMZN, while DeepAR reached 100% for AMGN and CVS
  • Notable case: For CVS, DeepAR significantly outperformed with 100% direction accuracy vs Fusion's 25%
Performance metrics comparison table
Forward Prediction Visualization

The 5-day forward predictions reveal interesting patterns for different companies:

amazon comparison plot cvs comparison plot

Both models predict a downward trend for Amazon, but the Fusion Layer tracks closer to actual values in early days before diverging. Models show opposing predictions for CVS: DeepAR correctly forecasts a downward movement while Fusion Layer incorrectly predicts an upward trend.

Key Takeaways

Our evaluation reveals that neither model consistently outperforms the other across all companies and metrics:

  • DeepAR strengths: More consistent prediction accuracy, particularly strong with healthcare stocks (ABT, AMGN, CVS)
  • Fusion Layer strengths: Better performance with tech companies (AMZN, GOOG), excellent direction accuracy for certain stocks
  • Practical application: A hybrid approach using both models with company-specific weighting could yield optimal results