William Bourgeois
White Papers

Research

Applied research at the intersection of machine learning and finance. Full PDFs are available for download.

Comparing Momentum and Support-Resistance Indicators

Rising Visibility Graphs, Linear Regression Channels, and DeMark TDST · William Bourgeois · McGill University, 2026

Price chart with a linear regression channel and visibility graph edges, on a crumpled paper against a blue background

Many portfolio managers rely on heuristic indicators to time entries and exits. One of the most widely used, Tom DeMark's Setup Trend (TDST), has no formal statistical foundation. This paper asks whether methods with stronger mathematical grounding produce better signals.

Three approaches are compared on 93 large-cap equities over 10 years of daily data: TDST (the practitioner heuristic), Linear Regression Channels (a classical statistical method), and Rising Visibility Graphs (a graph-theoretic method from network science). Each is evaluated on momentum signal quality (Information Coefficient, Forward Expectancy) and support-resistance accuracy.

No reliable momentum No method produces a reliable momentum signal at any tested horizon, on both benchmark-normalized and raw prices.
40% more accurate levels LR and RVG support-resistance levels are roughly 40% more accurate than TDST at anticipating realized price reversals.
Mean reversion, mostly LR support levels show weak but statistically significant predictive power, largely attributable to generic mean reversion.
Download PDF →

Predicting Credit Ratings Using Supervised Data Mining Models

William Bourgeois et al. · April 2024

Accurate credit rating prediction plays a pivotal role in risk assessment for financial institutions, investors, and businesses. This study compares supervised data mining models trained to predict company credit ratings from fundamentals.

Financial features (debt, equity, assets, market capitalization, liquidity, profitability, solvency, efficiency, growth, and volatility metrics) were collected via the Yahoo Finance API for companies on the NYSE, AMEX, and NASDAQ, and matched against S&P, Moody's, and Fitch ratings for the Russell 3000.

Five model families were evaluated: support vector machines, decision trees, logistic regression, random forests, and multilayer perceptron neural networks. The multilayer perceptron delivered the strongest performance after hyperparameter search.

Download PDF →
Model comparison chart: accuracy plus F1 for MLP, random forest, logistic regression, decision tree, and SVM