Prometheus

Setting Up Environment

Use Python 3.11

pip install -r requirements.txt

---

Project Overview

Prometheus focuses on analyzing relationships between securities and building predictive models for time-series data. This project leverages both statistical methods and machine learning models to uncover insights and make accurate forecasts.

---

Methods and Models

1. Statistical Analysis

• Multivariate Linear Regression: Evaluates relationships between multiple variables.
• Granger Causality Tests: Explores causal relationships between time-series.
• Cointegration Analysis: Tests long-term equilibrium relationships.
• Dynamic Time Warping (DTW): Measures similarity between time-series.

Initial (Deprecated) Predictive Models

• AutoARIMA: Statistical time-series forecasting with automated tuning.
• Deep Learning Models: Implements advanced models such as NHITS and NBEATS for forecasting.
• Baseline Methods: Includes benchmarks like linear regression and horizontal line predictions for comparison.

---

Machine Learning Architectures Attempted

- Encoder-Decoder Transformer

• Implements an encoder-decoder architecture with Transformer layers for sequence-to-sequence forecasting.
• Features:
  • Triple Positional Encoding: Combines embeddings for feature types, tickers, and time steps.
  • Supports multi-head attention mechanisms and layer normalizations.
  • Uses auxiliary losses to improve generalization and stability.

- StockBERT

• A specialized variant of BERT adapted for financial data.
• Features:
  • Multi-dimensional token embeddings for tickers, time indices, and continuous features.
  • Leverages transformer encoder layers for masked language modeling (MLM) tasks.
  • Integrates positional and temporal embeddings (e.g., yearly, monthly, and hourly).

- Somoformer

• A novel architecture that uses Transformer encoders tailored to handle sequential financial data.
• Features:
  • Incorporates triple positional encodings.
  • Adjusts output values using post-processing techniques, including mean-offset fixes.
  • Designed for sequence-to-sequence learning with flexible input-output sizes.

- Variational Autoencoder (VAE)

• Captures latent representations of time-series data for reconstruction and forecasting.
• Features:
  • Encoder-decoder design utilizing convolutional layers for spatial embeddings.
  • Supports both smaller and larger model versions, with optional Discrete Cosine Transform (DCT) encoding.
  • Implements KL divergence losses for latent regularization.

3. Statistical and Spatial-Temporal Deep Learning

Statistical methods like Dynamic Time Warping (DTW) and causal inference (e.g., Granger causality) were coupled with deep learning architectures for additional layers of interpretability and adaptability to financial time series.

Future Directions

• Statistical Analysis of dataset to identify explainable key relationships and patterns.
• Polynomial regression analysis of Stock Relationships to identify broad relationships between securities for further analysis.
• Further experimentation with hybrid statistical and machine learning pipelines (using Trees and Deep learning).
• Implementation of ensemble approaches combining multiple models.
• Updated Evaluation against additional datasets and real-world use cases.