Projects

Problem

Standard portfolio allocators do not internalise transaction costs, leading to high turnover and erosion of returns.

Approach

Portfolio allocation across HDFC, ICICI, and Kotak (2018-2023) formulated as a continuous-action MDP. Reward combines returns with a turnover penalty. PPO is trained against this reward and evaluated on returns, turnover, and drawdown profile.

Methodology

• Universe: HDFC, ICICI, Kotak daily data, 2018-2023.
• Action space: continuous allocation weights, softmax-constrained.
• Reward: returns net of turnover penalty.
• Trainer: PPO with standard actor-critic.
• Evaluation: portfolio evolution and drawdown curves alongside summary metrics.

Results

• 155.7% cumulative return over the test window.
• Sharpe ratio of 0.730.
• Turnover of 0.683, indicating learned cost awareness.
• Drawdown profile consistent with allocator behaviour rather than aggressive trading.

Problem

Per-bus load forecasts ignore network topology and treat the grid as independent time series.

Approach

A graph neural network over the IEEE-14 bus system carries information across nodes. Spatial structure is combined with temporal modelling. Compared against statistical and deep learning baselines through a modular PyTorch pipeline.

Methodology

• Graph: IEEE-14 bus topology with load measurements at each node.
• Model: spatio-temporal GNN with message passing and sequence modelling.
• Baselines: statistical and deep learning short-term forecasters.
• Engineering: modular PyTorch pipeline for reproducibility.

Results

• Outperformed statistical and deep learning baselines on short-term load forecasting.
• Captured spatio-temporal correlations not available to per-node models.