Research Assistant · IIT Kharagpur
Prototype
?
Stylized forecast shootout on a synthetic anomaly series — mirroring the real result: statistical models beat plain regression by ~10%, and the custom DVS-LSTM set the error minimum.
Highlights
- Comparative Framework: Conducted a comprehensive research study benchmarking 6 statistical, machine learning, and deep learning architectures for long-term climate anomaly forecasting across 50+ years of environmental historical records.
- Data Processing: Cleaned and analyzed over 500,000 telemetry records using the Average Ocean Temperature Anomaly Dataset.
- Model Training: Trained and hyperparameter-tuned ARIMA, Holt-Winters (HWELTS), Linear Regression, Deep Neural Network (DNN), and LSTM models via structured grid search optimizations.
- Novel Model Architecture: Co-designed a custom hybrid architecture—Deep Visibility Series with LSTM (DVS-LSTM)—which utilized visibility graph mappings to capture hidden temporal boundaries.
- Performance Benchmark: Evaluated models using MSE and MAE metrics. Found that while statistical models (ARIMA/HWELTS) reduced error metrics by ~10% compared to baseline regression, the custom DVS-LSTM out-performed all deep learning baselines by a margin of ~2–5%, establishing an error minimum.
Technologies
- ARIMA
- Holt-Winters
- LSTM
- DNN
- Grid Search
- DVS-LSTM