QuantumForge AI: The Ultimate Trading & Economic Assistant, Market Price Prediction To Deep Research.
November 25, 2025
AI Generated Image
QuantumForge AI is an enhanced stock prediction system— a comprehensive Python-based framework designed for stock price forecasting, technical analysis, and generative AI–driven insights, Deep Research Agent. It leverages deep learning (BiGRU with Attention and Conv1D layers), technical indicators, and integrates with Google's Gemini AI for automated investment reports and a conversational research agent.
The system supports real-time data fetching via yfinance, model training, evaluation, multi-day forecasting, and interactive querying for stocks, market news, or deep research topics.
Built for quantitative analysts, traders, and market enthusiasts, it provides:
- Predictive Modeling: Utilizes historical data to forecast stock prices with uncertainty intervals.
- Visualization: Dark-themed plots for distributions, correlations, test results, and forecasts.
- AI Augmentation: Generates professional-grade reports and handles natural language financial queries.
- Extensibility: Modular design enables easy fine-tuning or the addition of new indicators and models.
The project emphasizes reproducibility (seeded randomness), efficiency (mixed-precision training, caching), and usability (a conversational CLI loop).
As of November 25, 2025, the project uses up-to-date libraries and assumes access to a GPU for optimal performance.
At any time during the session, you can request instant stock quotes, 10-day price history tables, or full forecasting pipelines. For example: “stock GOOGL” → shows current price, volume, and recent performance table. “predict Amazon” → trains model, displays actual vs predicted chart, 4-day forecast with confidence interval, and saves a complete AI-generated report.
Jupyter Notebook Project Output Image
The EnhancedGRU(nn.Module) model is designed for deep temporal forecasting using sequence inputs of shape (batch, 40, ~18), consisting of OHLC features and technical indicators. It begins with multi-kernel Conv1D layers (kernels 3 and 5) that extract localized patterns and concatenate their outputs to expand the feature dimension. These enriched sequences pass into a bidirectional GRU (1 layer, 128 hidden units per direction, 256 total) that captures both forward and backward temporal dependencies. A scaled dot-product attention mechanism then weights GRU outputs to emphasize the most informative time steps. The regression head consists of a Linear layer (256 → 1), LayerNorm, and Dropout(0.15). The full forward pass follows Conv → GRU → Attention → Linear, using ReLU activations. The model is trained with Huber (SmoothL1) loss for robustness to outliers and automatically selects CUDA when available, otherwise defaults to CPU.
The Trainer(Config) class handles the full training workflow using the Adam optimizer (learning rate 5e-4) along with a ReduceLROnPlateau scheduler set with a patience of 10 epochs. Training is performed with a batch size of 64 for up to 50 epochs, using an early-stopping mechanism with a patience of 15 to prevent overfitting. Mixed-precision training (AMP) with GradScaler is enabled to improve both speed and memory efficiency. Throughout training, the system monitors training and validation loss, automatically saving the best-performing model to models/{ticker}_model.pth.
 |
 |
Jupyter Notebook Project Output Image
| Category | Feature Description |
|---|---|
| Data Pipeline | Downloads & preprocesses stock data with 15+ technical indicators (RSI, MACD, Bollinger Bands, ATR, EMAs, Momentum, Volatility, etc.) |
| Advanced Model | Bi-directional GRU with multi-scale Conv1D feature extraction + self-attention for deep temporal pattern learning |
| Training Enhancements | Early stopping, learning rate scheduling, gradient clipping, and mixed precision (AMP) for faster & stable convergence |
| Evaluation Metrics | RMSE, MAE, R², directional accuracy + backtesting on holdout data |
| Forecasting | Recursive multi-step predictions (default: 4 days) with confidence intervals |
| Visualizations | Feature histograms, correlation heatmaps, actual vs predicted plots, forecast charts |
| AI-Generated Reports | Gemini-powered Markdown reports with executive summaries, recommendations (e.g., Strong Buy), and key technical levels |
| Conversational Interface | Commands like "predict Apple", "deep research on AI stocks", or general queries like "latest on tech stocks?" |
| Caching & Persistence | Automatically saves data & trained models; supports resume training |
| Research Agent | Deep research using DuckDuckGo/Wikipedia with fact-checking & structured Markdown output |
The DataHandler(Config) class manages the complete data pipeline, beginning with downloading OHLCV price data from yfinance (default range: January 1, 2022 to the current date). It computes a rich set of technical indicators, including SMA-10, RSI-14, MACD-12/26, Bollinger Bands-20, ATR-14, price change, log-close, volatility windows, momentum shifts, and multiple EMAs. After indicator generation, the handler cleans the dataset by resolving NaN/inf values using forward-fill, mean imputation, and zero-filling. All features are scaled using MinMaxScaler, and the system generates sliding-window sequences (default: 40-day windows). The dataset is then split into an 80/20 train-test ratio and cached as pickles for faster reuse. The class outputs scaled training and testing sequences—(X_train, y_train) and (X_test, y_test)—along with the fully scaled dataset, and also provides an inverse transformation utility to convert predictions back into real USD prices.
The system uses Gemini with the configuration gemini-2.0-flash (tokens=1600, temp=0.7) and supports a report generator that produces structured Markdown outputs containing sections such as Summary, Performance, Forecast, Risk, Recommendation, and Targets. Each report is saved following the pattern reports/APPLE_REPORT.md. For general interactions, the assistant switches to a "Goldman Sachs strategist" persona to handle non-stock queries in a more conversational and analytical tone.
The Report Generation Agent automates the creation of complete, publication-ready investment reports immediately after each stock prediction. Once the deep learning pipeline — featuring Conv1D, BiGRU, and attention mechanisms — completes training and produces a 4-day price forecast with confidence intervals, the system automatically saves high-resolution dark-themed plots, including actual vs. predicted charts, forecast charts with shaded uncertainty, and feature correlation heatmaps. It compiles key numerical results such as RMSE, MAE, R², directional accuracy, predicted prices, and support/resistance levels, then sends all the data, insights, and embedded plots directly to Google Gemini. Gemini generates a professional, beautifully formatted Markdown report, typically 600-700 words, that includes an executive summary, model performance and accuracy metrics, short-term price forecasts with exact targets, embedded comparison graphs, technical analysis highlights, risk assessment, clear trading recommendations , providing a comprehensive, ready-to-use report for traders and investors.
The research agent provides multi-depth analysis through functions like deep_research(topic, depth='normal') and agent(user_input), drawing from DuckDuckGo's JSON API (10 results) and Wikipedia summaries, with ranking based on snippet length. It aggregates scraped content and sends it through a Gemini prompt to produce a structured ten-section research report with citations, followed by a secondary Gemini fact-checking pass. Users can control depth using commands such as "research X" for normal depth, "quick research X" for shallow output, and "deep research X" for a more detailed, expert-level analysis.
For evaluation, the trainer computes RMSE, MAE, and R² using inverse-transformed predictions to return real USD price metrics, along with Directional Accuracy, which measures the percentage of correctly predicted upward or downward price movements. Visual evaluation includes a dark-themed line plot comparing actual vs. predicted prices, using green and cyan lines for clear contrast and readability.
 |
 |
Jupyter Notebook Project Output Image
Clone/Setup Environment:textgit clone https://github.com/say217/QuantumForge-AI.git
cd enhanced-stock-predictor
python -m venv env
source env/bin/activate # Linux/Mac
# or env\Scripts\activate # Windowspip install -r requirements.txt.Go to AI Studio Open your browser and visit the AI Studio website. Log in / Sign up Use your email or GitHub/Google account to sign in. Create a Free-Tier API Key Find the API Keys section (usually under Settings or Developer menu). Click Create New API Key. Select Free Tier and confirm. Copy the API Key Once the key is created, click Copy to copy it to your clipboard.
Set GOOGLE_API_KEY env var (for Gemini): export GOOGLE_API_KEY="your_key".
yfinance handles Yahoo Finance implicitly (no key needed)
The image below shows a heat map of the correlation between the feature data, Price-based features such as Close, High, Low, SMA, EMA, and Log-Close show extremely high correlations (0.97–1.00), which is expected because they are all derived from the same underlying price series. Bollinger Bands (Upper and Lower) also correlate strongly with price (0.95–0.99), indicating that they largely track price movements and may offer limited additional information unless paired with volatility or trend indicators. In contrast, RSI, MACD, and Momentum features display only moderate correlations (generally 0.17–0.70), suggesting they capture shifts in trend and momentum that could add predictive value. Volatility features (10, 20, 50-period) are among the least correlated, making them more independent and potentially useful for diversifying the feature set. Also the below image of the LLM model’s output, demonstrating that the language model (Gemini-2.0-Flash) behaves like an economic specialist and provides research-based insights. However, users should always double-check the information and avoid making decisions solely based on AI. This is only a prototype disclaimer.”
 |
 |
Jupyter Notebook Project Output Image
The Enhanced Stock Predictor helps by automating and streamlining tasks that are normally time-consuming, complex, or require expert knowledge. It solves key problems for traders and investors, such as:
-
Eliminates manual analysis: Instead of spending hours calculating technical indicators, scanning charts, and researching news, the system provides forecasts, signals, and AI-generated insights instantly.
-
Reduces information overload: Traders no longer need to sift through dozens of news sources or reports; the AI synthesizes relevant data and presents concise, actionable information.
-
Improves decision-making: By combining deep learning forecasts with technical indicators and confidence measures, it helps users time entries and exits more accurately.
-
Supports multi-stock monitoring: Investors managing multiple positions can get live updates and forecasts for several stocks at once, reducing the risk of missing opportunities.
-
Speeds research and strategy development: With automated, cited, and fact-checked reports, users can quickly explore sectors, compare stocks, and backtest strategies without extensive manual work.
-
Bridges expertise gaps: Beginners or part-time traders gain access to professional-level insights and quantitative analysis that would normally require years of experience or expensive tools. In short, it turns complex, repetitive, and data-heavy trading tasks into a fast, reliable, and AI-driven workflow, helping traders save time, reduce errors, and make more informed investment decisions.
Desclimer REGULATORY NOTICE This tool is not registered with the SEC, FINRA, FCA, or any other financial regulatory authority.
By using QuantumForge AI, you acknowledge and accept this disclaimer in full.
This is a student-based project created to demonstrate my skills and ideas. It is only a prototype and should not be used for real financial decision-making.
(November 25, 2025)
NOT PROFESSIONAL FINANCIAL ADVICE
QuantumForge AI is an experimental, educational, and research-oriented tool created for entertainment, learning, and quantitative exploration purposes only.
The predictions, forecasts, investment reports, recommendations (Strong Buy / Buy / Hold / Sell / Strong Sell), price targets, and any opinions generated by the system — whether from the neural network model or Google’s Gemini AI — do NOT constitute financial, investment, or trading advice.
USE AT YOUR OWN RISK
You alone are responsible for any trading, investment, or financial decisions you make.
The authors, contributors, and distributors of QuantumForge AI accept zero liability for any financial losses, missed gains, emotional distress, or any other consequences arising from using this software.
REGULATORY NOTICE
This tool is not registered with the SEC, FINRA, FCA, or any other financial regulatory authority.
By using QuantumForge AI, you acknowledge and accept this disclaimer in full.
This is a student-based project created to demonstrate my skills and ideas. It is only a prototype and should not be used for real financial decision-making.