📈 Autonomous Stock Trading System

An end-to-end pipeline combining NLP sentiment analysis, LSTM price forecasting, and Deep Reinforcement Learning to train an autonomous trading agent - validated on Tesla (TSLA) stock data.

---

🗺️ System Overview

The pipeline runs in four stages across five Jupyter notebooks:

🗞️  Raw Data (tweets + OHLCV prices)
         │
         ▼
🧠  [1] Financial_Sentiment_Extraction_System.ipynb
         Extracts per-day sentiment scores using FinBERT
         │
         ▼
🔮  [2] predicting_close_LSTM.ipynb
         Trains a 3-layer LSTM to predict next-day closing prices
         │
         ▼
🔧  [3] Finalising_predicting_close_LSTM.ipynb
         Filters dates and assembles final RL training datasets
         │
         ▼
🤖  [4a] gym_anytrading_with_technical_indicators_and_sentiment.ipynb
         Trains the A2C trading agent with the full feature set

🔬  [4b] Abelation Test.ipynb
         Repeats training across feature subsets to isolate which
         inputs drive performance

---

📁 Repository Structure

algo-trading-system/
├── 📂 Code/
│   ├── 1. Sentiment Extraction/
│   │   └── 🧠 Financial_Sentiment_Extraction_System.ipynb
│   ├── 2. Close Price Prediction/
│   │   ├── 🔮 predicting_close_LSTM.ipynb
│   │   └── 🔧 Finalising_predicting_close_LSTM.ipynb
│   └── 3. RL Logic/
│       ├── 🤖 gym_anytrading_with_technical_indicators_and_sentiment.ipynb
│       └── 🔬 Abelation Test.ipynb
└── 📂 Data/
    ├── Tweets data/
    │   ├── tweets.csv
    │   └── processed_tweets.csv
    └── TESLA data/
        ├── 1. Raw data/
        │   └── PAPER_TSLA_data.csv
        ├── 2. Processed data/
        │   ├── processed_PAPER_TSLA_data.csv
        │   └── processed_PAPER_TSLA_data_probabilities.csv
        ├── 3. Normalised data/
        │   ├── processed_normalised_PAPER_TSLA_data.csv
        │   └── processed_normalised_PAPER_TSLA_data_probabilities.csv
        ├── 4. LSTM data/
        │   ├── processed_normalised_PAPER_TSLA_data_LSTM.csv
        │   └── processed_normalised_PAPER_TSLA_data_probabilities_LSTM.csv
        └── 5. RL training data/
            ├── TESLA RL training data - 1 Sentiment 12.2024.csv
            └── TESLA RL training data - 3 Sentiment 12.2024.csv

---

📦 Required Input Data

🗞️ 1. News / tweet headlines CSV

In our research we used this Kaggle dataset of Tesla-related tweets.

Required columns:

• Date - publication date of the headline
• Tweet - headline or news text

💹 2. Stock price CSV

We used Yahoo Finance to download TSLA OHLCV data (2013–2020).

Required columns:

• Date - trading day
• Open, High, Low, Close, Adj Close, Volume

---

🚀 Usage: Step-by-Step

🧠 Step 1 - Extract Sentiment

Notebook: Code/1. Sentiment Extraction/Financial_Sentiment_Extraction_System.ipynb

Cleans raw tweets, runs them through FinBERT (a pre-trained financial BERT model), and aggregates daily sentiment scores. Days with no news receive a neutral score of 0.

Configure the ADD_ONE_SENTIMENT_COLUMN flag to choose the output format:

• ✅ True → a single Sentiment column with values {-1, 0, 1}
• 🔢 False → three probability columns: Positive Sentiment, Negative Sentiment, Neutral Sentiment

In our research, the ADD_ONE_SENTIMENT_COLUMN flag is set to True.

The notebook also Z-score normalises the financial features (Open, High, Low, Close, Adj Close, Volume).

📤 Outputs:

• Data/Tweets data/processed_tweets.csv
• Data/TESLA data/2. Processed data/processed_PAPER_TSLA_data.csv
• Data/TESLA data/3. Normalised data/processed_normalised_PAPER_TSLA_data.csv

---

🔮 Step 2 - Predict Closing Prices

Notebook: Code/2. Close Price Prediction/predicting_close_LSTM.ipynb

Trains a 3-layer LSTM network on 5 features (Open, High, Low, Close, Volume) using an 80-day sliding window. Predictions are inverse-transformed back to the original price scale and saved in a new Predicted_Close column.

⚙️ Key Hyperparameters:

Parameter	Value
🪟 Window size	80 days
🏗️ LSTM layers	3
🧩 Hidden units	150
💧 Dropout	60%
✂️ Train/val split	95% / 5%
⚡ Optimizer	Adam (lr=0.001)
🔁 Max epochs	50 (early stopping, patience=10)

📤 Output: Data/TESLA data/4. LSTM data/processed_normalised_PAPER_TSLA_data_LSTM.csv

---

🔧 Step 3 - Finalise RL Training Data

Notebook: Code/2. Close Price Prediction/Finalising_predicting_close_LSTM.ipynb

Filters the LSTM-augmented data to dates from 2014-01-01 onwards and produces two versioned CSV files - one for each sentiment format - ready for RL training.

📤 Outputs:

• Data/TESLA data/5. RL training data/TESLA RL training data - 1 Sentiment MM.YYYY.csv
• Data/TESLA data/5. RL training data/TESLA RL training data - 3 Sentiment MM.YYYY.csv

---

🤖 Step 4a - Train the RL Trading Agent

Notebook: Code/3. RL Logic/gym_anytrading_with_technical_indicators_and_sentiment.ipynb

Trains an A2C (Advantage Actor-Critic) agent inside a custom gym-anytrading environment while setting a trading WINDOW_SIZE. The agent learns a long/short trading policy over ~1,258 training days (up to 2019-01-01) and is evaluated on ~252 test days.

In our research, WINDOW_SIZE was tested across {10, 15, 20, 25}.

🏆 Performance is reported as: Calmar Ratio · Total Reward · Total Profit

---

🔬 Step 4b - Ablation Study

Notebook: Code/3. RL Logic/Abelation Test.ipynb

Runs the same A2C training loop for the selected FEATURE_STATE to measure the contribution to final trading performance. The FEATURE_STATE parameter controls which signals the agent observes:

🏷️ FeatureState	📡 Signals
BasicMode	Close price only
WithTechInd	Close + SMA, RSI, MOM, EMA, AROONOSC
WithPredict	Above + Predicted_Close
WithLag	Above + lag features
FullMode	Above + Sentiment

---

🛠️ Key Technologies

Component	Library / Model
🧠 Sentiment analysis
🔮 Price prediction
🎮 RL environment
🤖 RL algorithm
📊 Technical indicators
🔢 Data / ML utilities