Investment Advisory & Portfolio Management System
The Architecture
Financial decision-making is one of the highest-stakes domains for software to touch. This project started as a question: how close can a self-built system get to the kind of predictive tooling that institutional investors pay seven figures for?
The system is built around two distinct ML model classes, each chosen for a specific forecasting task. ARIMA (AutoRegressive Integrated Moving Average) handles time-series prediction for assets with strong temporal autocorrelation — stock prices and gold futures, where historical patterns carry statistically meaningful forward signal. Random Forest handles risk classification — a supervised learning approach that categorises portfolio positions by risk profile based on multi-feature input (volatility, sector, market cap, historical drawdown).
Real-time market data is sourced via the Yahoo Finance API, feeding both the live dashboard and the model inference pipeline. This means predictions aren't made against stale data — every forecast runs against the most recent available price history. Gold and real estate forecasting extends the model surface beyond equities, reflecting a realistic multi-asset portfolio view.
The frontend is intentionally lightweight — Flask templates with Plotly for interactive charting and Matplotlib for static visualisations. No JavaScript framework overhead; the complexity budget was reserved entirely for the ML layer. The result is a system where the intelligence is in the models, not the UI.
Risk assessment and classification sits at the core of the advisory layer — each portfolio position receives a risk tier, and the recommendation engine surfaces rebalancing suggestions based on the aggregate risk profile. SQLite handles local persistence for development with a PostgreSQL migration path documented for production use.
The repo is structured in three clean layers — /frontend, /backend, /docs — with an 18-commit history reflecting iterative model improvement rather than a single dump. This is experimental, but it's engineered experimentally.
Strategic Methodology
Model-first development — both ML models were trained, evaluated, and validated against historical Yahoo Finance data before any web interface was built. The Flask app was treated as a presentation layer for already-working intelligence, not a framework to build intelligence around. Documentation lives in /docs as a first-class output alongside the code.
Engineering Challenges
- Choosing the right model per asset class — ARIMA requires stationary time-series data, which meant applying differencing and ADF tests to each asset's price history before fitting. Random Forest required careful feature engineering to avoid lookahead bias in the training set.
- Integrating real-time Yahoo Finance API data into the model inference pipeline without introducing latency that degrades the user experience — solved with background data caching and incremental model updates rather than full retraining on each request.
- Extending forecasting beyond equities to gold and real estate — assets with fundamentally different volatility profiles, seasonal patterns, and data availability required separate preprocessing pipelines and model hyperparameter tuning per asset class.
- Building a risk classification system that's meaningful without being overconfident — the Random Forest outputs probability distributions per risk tier rather than hard labels, surfacing the model's uncertainty to the user rather than hiding it behind a single verdict.
- Visualising multi-dimensional financial data (price history, predicted trajectory, confidence intervals, portfolio allocation) in a web interface without overwhelming non-technical users — Plotly's interactive charts with progressive disclosure (hover for detail, zoom for precision) solved this cleanly.
Project Impact
"Where machine learning meets market reality — an AI-powered advisory system that applies ARIMA and Random Forest models to real-world financial data, because predicting the future is a problem worth engineering."
Core Arsenal
Technical Log.
A high-fidelity breakdown of the build's architectural achievements and performance markers.
Synthesis
"An AI investment advisory web app using ARIMA for time-series stock and gold forecasting, Random Forest for risk classification, Yahoo Finance API for real-time data and Plotly for interactive portfolio visualisation. Built as personal ML exploration"
Hard Evidence
Demonstrates ML engineering depth — two distinct model classes (ARIMA for time-series, Random Forest for classification) applied to a real-world financial domain with real market data.
Extends forecasting surface to three asset classes — equities, gold, and real estate — each with tailored preprocessing and model configuration.
Risk classification outputs probability distributions, not hard labels — a deliberate design decision that surfaces model uncertainty rather than false precision.
Proves Python versatility across three project types: FastAPI REST APIs (Curator), Flask ML apps (this), and a professional TypeScript stack — a genuinely polyglot engineering profile.
Real-time Yahoo Finance integration means predictions run on live data — not a toy trained on a static CSV.