A machine learning framework that quantifies the relative influence of five key factor categories on consumer demand — going beyond prediction to answer which factors matter most, and by how much.

Most demand forecasting systems tell you what demand will be. This project answers why — by systematically measuring the contribution of weather, economic conditions, social media trends, financial markets, and calendar effects to demand patterns using a multi-model interpretable ML approach.

Built as an academic study project at BITS Pilani (Mathematics Dept.), this framework is designed to be industry-adaptable and business-interpretable.

Businesses monitor many demand signals — weather, sentiment, macro indicators — but lack a quantitative basis for prioritizing which factors deserve investment. This project answers:

• How much does each factor category influence demand?
• Which signals should forecasting models prioritize?
• How can ML provide interpretable, quantified insights — not just predictions?

Demand-Forecasting/
│
├── Demand_Forecasting.ipynb      # Full pipeline: data → features → models → analysis
│
├── reports/
│   └── report.pdf              # Full project report
│
├── presentation/
│   └── presentation.pdf        # Project presentation slides
│
├── LICENSE                     # All rights reserved
└── README.md

Five factor categories integrated from real APIs (with synthetic fallback):

Pipeline auto-falls back to calibrated synthetic generation if API keys are unavailable.

47 features engineered across 6 groups, including:

• 7-day and 30-day rolling averages for weather, financial, and social signals
• Lag features (1-day, 7-day) for time-series awareness
• Interaction features: heat index, sentiment × shopping trend, GDP × interest rate
• Full temporal decomposition: quarter, week-of-year, day-of-year

All features standardized via StandardScaler (RF/SHAP) or MinMaxScaler (LSTM).

Synthetic demand uses controlled, known coefficients enabling ground-truth validation:

base_demand         = 1,500 units
GDP Growth          = +40 units/point
Inflation           = −25 units/point
Unemployment        = −20 units/point
Consumer Confidence = +4 units/point above 50
Shopping Trend      = +6 units/trend point
Smartphone Trend    = +4 units/trend point
National Holiday    = +500 units
Other Holiday       = +250 units
Weekend (Sat/Sun)   = +200 / +150 units
Festival Months     = +300 units (Sep–Nov)
Noise               = ±75 units (Gaussian, seed=42)

LSTM outperforms Random Forest across all metrics, capturing temporal dependencies that tree-based models miss. Both models provide a solid baseline for multi-factor demand modelling on synthetic data.

Factor importance is aggregated by summing feature-level scores within each group. Measured using both RF default importance and SHAP mean absolute values:

Consistent finding across both methods: Social media trends and temporal patterns are the dominant demand drivers, followed by weather and financial signals — challenging traditional macro-economic-centric forecasting assumptions.

Full pie chart breakdowns with precise percentages are available in the notebook outputs and project report.

# Clone the repository
git clone https://github.com/Haryaksh1/Demand-Forecasting.git
cd Demand-Forecasting

# Install dependencies
pip install pandas numpy scikit-learn shap tensorflow matplotlib seaborn holidays requests

# Add your API keys in the notebook (optional — synthetic fallback works without them)
# OPENWEATHER_API_KEY = "your_key_here"
# ALPHAVANTAGE_API_KEY = "your_key_here"

# Run
jupyter notebook m3sop_final_code.ipynb

• Synthetic data may not capture all real-world demand complexities
• Factor importance assumes temporal stationarity
• Calibrated to Indian market conditions (Delhi weather base, India public holidays)
• SHAP for LSTM skipped due to TensorFlow/Keras version compatibility
• 730-day observation window

• Real-world validation with actual retail or e-commerce data
• Time-varying (rolling window) factor importance
• Causal inference beyond correlation
• LSTM SHAP integration with compatible explainer
• Industry-specific calibration

All Rights Reserved.

This repository is shared for viewing and academic reference only. No part of this code, methodology, or documentation may be copied, modified, redistributed, or used in any form — commercial or otherwise — without explicit written permission from the author.

© 2025 Haryaksh Manuh Bhardwaj. All rights reserved.

1. Lundberg & Lee (2017) — A Unified Approach to Interpreting Model Predictions. NeurIPS.
2. Ke et al. (2017) — LightGBM: A Highly Efficient Gradient Boosting Decision Tree. NeurIPS.
3. Chen et al. (2025) — Supply Chain Demand Forecasting based on Multi-Time Scale Data Fusion. Computers & Industrial Engineering.
4. Tadayonrad & Ndiaye (2023) — A new KPI model for demand forecasting in inventory management. Supply Chain Analytics.
5. Zheng & Casari (2018) — Feature Engineering for Machine Learning. O'Reilly Media.
6. Choi & Varian (2012) — Predicting the Present with Google Trends. Economic Record.
7. MIT Sloan — What is synthetic data and how can it help you competitively?