Team Dynamo

FidelFolio Market Cap Growth Forecasting

Project Overview

This repository presents our solution to the FidelFolio Deep Learning Challenge, which involves predicting market capitalization growth across multiple time horizons for Indian companies. Using deep learning, we aim to model the nonlinear relationships between financial indicators and future market performance.

---

Objective

• Predict market cap growth across:
  • Short-term (1-Year): Target 1
  • Medium-term (2-Year): Target 2
  • Long-term (3-Year): Target 3
• Compare deep learning models to identify the most accurate and robust architecture.

---

Dataset

• Instances: Company-Year combinations
• Features: 28 fundamental financial indicators (Feature1 to Feature28)
• Targets:
  • Target 1 (1Y growth)
  • Target 2 (2Y growth)
  • Target 3 (3Y growth)

Preprocessing Steps

• Missing values imputed using company-wise and global means.
• Winsorization applied to cap outliers.
• Standardization performed using StandardScaler.

---

Models Implemented

1. Multilayer Perceptron (MLP)

• Basic feedforward neural network with dropout and ReLU activation.
• Trained for 1000 epochs.
• MLP is not able to capture the complex Time Series Patterns in the Dataset.

Performance (RMSE):

• Target 1: 106.6246
• Target 2: 227.6656
• Target 3: 370.3143

---

2. LSTM (Vanilla)

• Standard sequence model using final hidden state.
• Trained for 2000 epochs.

Performance (RMSE):

• Target 1: 22.7029
• Target 2: 56.9746
• Target 3: 186.7859

---

3. LSTM with Attention

• Incorporates soft attention over time for feature weighting.
• Trained for 3000 epochs.

Performance (RMSE):

• Target 1: 19.4405
• Target 2: 47.9110
• Target 3: 175.1124

---

4. Transformer Encoder

• Uses positional encoding and multi-head self-attention.
• Trained for 1200 epochs.

Performance (RMSE):

• Target 1: 30.9814
• Target 2: 38.7712
• Target 3: 112.2152

---

5. DeepTCN (Temporal Convolutional Network)

• Causal convolutions with dilation to capture temporal dependencies.
• Results to be updated.

---

6. N-BEATS (Neural Basis Expansion)

• Forecasting-specific architecture designed for interpretability.
• Results to be updated.

---

7. DeepAR

• Autoregressive model trained on probabilistic distributions.
• Results to be updated.

---

8. Multi-horizon TCN

• TCN variant tailored for multi-horizon forecasting.
• Trained for 2000 epochs.

Performance (RMSE):

• Target 1: 28.5087
• Target 2: 34.4607
• Target 3: 42.4538

---

Model Comparison

Model	Attention	RMSE T1	RMSE T2	RMSE T3
MLP	No	106.6246	227.6656	370.3143
LSTM	No	22.7029	56.9746	186.7859
LSTM + Attention	Yes	19.4405	47.9110	175.1124
Transformer	Yes	30.9814	38.7712	112.2152
DeepTCN	Yes	Pending	Pending	Pending
N-BEATS	No	Pending	Pending	Pending
DeepAR	Yes	Pending	Pending	Pending
Multi-horizon TCN	Yes	28.5087	34.4607	42.4538

---

Multi-Horizon TCN is performing best among all.