SignalMatrix AI — Multi-Step Market Forecasting System

AI-Powered 10-Day NIFTY 50 Stock Price Forecasting & Strategy Evaluation Platform

---

Problem Statement

Retail investors and analysts often lack accessible tools to forecast short-term market trends from raw price data. SignalMatrix AI addresses this by providing a browser-based, no-code interface that takes NSE stock CSV exports, applies a trained LSTM model, and returns a 10-day forward price forecast with validation metrics — all without requiring any machine learning expertise.

---

Live Demo(Linkedin post)

-> Click here to view the live working & project's url

---

Pipeline

NSE India CSV Export (OHLC Daily Data)
        │
        ▼
Data Preprocessing
  Column normalization · Type casting · Date sorting · NaN removal
        │
        ▼
Feature Engineering (13 Features)
  MA_20 · EMA_20 · RSI · MACD · Signal Line · Bollinger Bands (Upper/Middle/Lower)
        │
        ▼
MinMaxScaler Normalization  (scaler.pkl)
        │
        ▼
60-Day Lookback Window  →  LSTM Seq2Seq Model  (lstm_model.keras)
  Layer 1: LSTM(128) + Dropout
  Layer 2: LSTM(64)  + Dropout
  Dense Head → 10-Step Output
        │
        ▼
Inverse Transform  →  Additive Bias Correction  →  EWM Smoothing
        │
        ▼
Plotly Chart  +  Validation Metrics (MAE / RMSE / Directional Accuracy)
        │
        ▼
Streamlit App  (app.py)

---

Key Features

• 10-Day Multi-Step Forecast — predicts a full 10-trading-day sequence, not just next-day
• Validation Mode — compare AI forecast against real data; computes MAE, RMSE, and Directional Accuracy
• 13 Engineered Features — OHLC + MA, EMA, RSI, MACD, Signal Line, Bollinger Bands
• Stock Symbol Auto-Detection — reads from CSV Symbol column or extracts from filename
• Additive Bias Correction + EWM Smoothing — corrects scale drift between last known close and first forecast step
• NSE-Compatible CSV Format — handles NSE's default column names (Open Price, High Price, etc.) automatically
• Docker-Ready — Dockerfile + .dockerignore included for containerised deployment
• Interactive Plotly Charts — Historical (blue) · Forecast (red) · Validation Actual (orange)

---

Model Architecture

Input Shape: (batch, 60, 13)
        │
        ▼
LSTM(128, return_sequences=True) → Dropout
        │
LSTM(64,  return_sequences=False) → Dropout
        │
Dense(64) → Dense(32) → Dense(10)
        │
Output Shape: (batch, 10)   ← 10-day price forecast

Parameter	Value
Architecture	Stacked LSTM (Seq2Seq)
LSTM Units	128 → 64
Lookback Window	60 trading days
Forecast Horizon	10 trading days
Total Parameters	124,522
Regularization	Dropout
Output Activation	Linear

---

Training Performance

Metric	Value
RMSE	135.78
MAE	58.60
MAPE	14.23%
R² Score	0.9766

Training Visualization — Actual vs Predicted (Full Dataset)

Last 100 Days Prediction Behavior

---

Dataset

Source: NSE India — Equity Security Data

Split	Rows (approx.)	Description
Training	~1,75,000	5 years of NIFTY 50 combined stock data (all 50 symbols)
Validation	~45,000	Held-out combined NIFTY 50 data

Column	Description
Date	Trading date
Open	Opening price of the session
High	Intraday high
Low	Intraday low
Close	Session closing price ← prediction target
Symbol	NIFTY 50 stock ticker (encoded via stock_mapping.pkl)

Note: NSE exports use column names like Open Price, High Price, etc. The app renames these automatically.

---

Feature Engineering

The model uses 13 features derived from raw OHLC price data:

Feature	Description
Symbol	Encoded stock identifier (from stock_mapping.pkl)
Open	Opening price
High	Intraday high
Low	Intraday low
Close	Closing price — prediction target
MA_20	20-day Simple Moving Average — medium-term trend
EMA_20	20-day Exponential Moving Average — recent-price weighted
RSI	Relative Strength Index (14-day) — overbought/oversold signal
MACD	EMA(12) − EMA(26) — momentum shift indicator
Signal_Line	9-day EMA of MACD — crossover detection
BB_Middle	20-day rolling mean — Bollinger Band centre
BB_Upper	Middle + 2σ — volatility resistance level
BB_Lower	Middle − 2σ — volatility support level

---

Chart Legend

Line Colour	Meaning
🔵 Blue	Last 60 days of historical close prices
🔴 Red	AI 10-day forecast
🟠 Orange	Validation Actual (only shown when Validation Mode is enabled)

---

Project Structure

SignalMatrix-AI/
├── app.py                      # Streamlit frontend + inference pipeline
├── lstm_model.keras            # Trained LSTM model weights
├── scaler.pkl                  # Fitted MinMaxScaler
├── stock_mapping.pkl           # NIFTY 50 symbol → integer encoding map
├── NIFTY50_all.csv             # Full training dataset (NSE export)
├── model_architecture.png      # LSTM architecture diagram
├── training_actual_vs_pred.png # Full dataset actual vs predicted chart
├── last_100_days.png           # Last 100 days forecast behaviour plot
├── logo.png                    # App logo (used in Streamlit header)
├── requirements.txt            # Python dependencies
├── Dockerfile                  # Container build file
└── .dockerignore               # Docker ignore rules

---

Quickstart

Local (Python)

# Clone repository
git clone https://github.com/Akshbhimani08/SignalMatrix-AI.git
cd SignalMatrix-AI

# Install dependencies
pip install -r requirements.txt

# Run app
streamlit run app.py

Docker

docker build -t signalmatrix-ai .
docker run -p 8501:8501 signalmatrix-ai

Then open http://localhost:8501 in your browser.

---

How to Use

Step 1 — Select Mode

Mode	Description
🚀 Future Forecast	Upload stock CSV → get 10-day price forecast
📊 Why Choose This Model?	View training metrics, architecture, and feature descriptions

Step 2 — Upload NSE Stock Data

• Download from NSE India → Equity Security → Historical Data
• File must be .csv or .xlsx
• Must contain: Date, Open / Open Price, High / High Price, Low / Low Price, Close / Close Price
• Minimum 80 rows required (≈ 4 months of trading data)

Step 3 — Generate Forecast

• Click Run to generate the 10-day forecast
• Optionally enable Validation Mode (requires 80+ rows) to compare forecast vs actual prices

Step 4 — Interpret Results

• Red line = AI Forecast
• Orange line = Validation Actual (if enabled)
• Blue line = Historical Data
• Metrics: MAE · RMSE · Directional Accuracy

---

Validation Metrics Explained

Metric	Formula	Interpretation
MAE	mean(|actual − predicted|)	Average absolute error in ₹
RMSE	√mean((actual − predicted)²)	Penalises large errors more heavily
Directional Accuracy	% correct trend direction (up/down)	How often the model gets the movement direction right

---

Important Notes

• Model trained on NIFTY 50 index stocks only — accuracy on mid/small-cap stocks outside the index is not guaranteed
• Uses a 60-day lookback window — upload at least 4 months of data for best results
• Additive bias correction is applied post-inference to anchor the first predicted point to the last known close price
• EWM smoothing (span=3) is applied to reduce step-to-step prediction noise
• This is an educational project — not intended for live trading or financial advice

---

References

• Hochreiter, S. & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation. doi:10.1162/neco.1997.9.8.1735
• Sutskever, I. et al. (2014). Sequence to Sequence Learning with Neural Networks. arXiv:1409.3215
• NSE India — Historical Equity Data
• Streamlit — streamlit.io
• Plotly Python — plotly.com/python

---

Citation

@misc{signalmatrix2025,
  author       = {Aksh Bhimani},
  title        = {SignalMatrix AI: Multi-Step LSTM Forecasting for NIFTY 50 Stocks},
  year         = {2025},
  howpublished = {\url{https://github.com/Akshbhimani08/SignalMatrix-AI}},
  note         = {Streamlit · TensorFlow · LSTM Seq2Seq · NSE India Data}
}

---

Author

Aksh Bhimani

---

Made for learning financial AI and time-series forecasting.