Forecasting USD PKR exchange rates using Urdu social media sentiment

Area of Science:

• Computational Finance
• Natural Language Processing
• Time Series Analysis

Background:

• Exchange rate forecasting traditionally relies on historical price data.
• Public sentiment, especially during instability, is an underutilized signal.
• The USD/PKR currency pair is crucial for Pakistan's economy.

Purpose of the Study:

• To develop a hybrid forecasting framework for the USD/PKR exchange rate.
• To integrate historical exchange rate data with sentiment analysis from Urdu tweets.
• To evaluate the effectiveness of different sentiment analysis models and forecasting architectures.

Main Methods:

• Collected 172,002 tweets from X (formerly Twitter) using trending hashtags.
• Filtered and processed 45,048 relevant Urdu-language tweets for sentiment analysis.
• Compared sentiment analysis models: Gemini 1.5 Flash, GPT-3.5 Turbo, GPT-4o, and XGBoost.
• Integrated sentiment scores with USD/PKR exchange rate data.
• Employed Long Short-Term Memory (LSTM), Xtreme Gradient Boosting (XGBoost), and a hybrid LSTM+XGBoost model for forecasting.

Main Results:

• The hybrid LSTM+XGBoost model incorporating GPT-4o sentiment scores achieved the best performance.
• Achieved a Root Mean Squared Error (RMSE) of 0.0831 and Mean Absolute Percentage Error (MAPE) of 0.03%.
• Outperformed the LSTM baseline model trained solely on historical data.

Conclusions:

• Urdu social media sentiment is a valuable predictor for exchange rate movements.
• Hybrid forecasting models are superior to standalone models for leveraging social media sentiment.
• The proposed framework offers improved accuracy for USD/PKR exchange rate forecasting.