Where Data Science
Meets Domain Expertise

Commodity Price Forecasting

Using deep learning architectures including LSTM networks and Transformer models to forecast commodity market prices across multiple time horizons. The models incorporate multi-variate time-series data and custom attention mechanisms for improved prediction accuracy.

Key Techniques

Approach

• • Multi-horizon forecasting with custom neural architectures
• • Feature engineering from macroeconomic and market indicators
• • Ensemble methods combining statistical and DL models
• • Backtesting framework for strategy evaluation

Retirement Decumulation: A Three-Paper Series

Ruin probability across withdrawal strategies at varying initial wealth levels

A three-paper program at Canada’s Financial Wellness Lab treating retirement decumulation as a stochastic control problem in an incomplete market with jumps, health shocks, taxes, and fees. Joint work with Matt Davison; all three papers are under review.

Key Techniques

Research Focus

• • Paper 1 (Merton-informed policies): a drift-penalized proportional rule cuts ruin probability by 17 to 22 percentage points versus the 4% Rule across $300k to $3M, with higher expected real income
• • Paper 2 (dynamic Bellman allocation with normalized utility): lower ruin and higher efficiency than the best static rule across $200k to $1M
• • Paper 3 (smile-optimal target-tracking): fits the empirical U-shaped retirement spending smile (high early, a trough near 80, a late-life rise)
• • Core research at Canada’s Financial Wellness Lab

Hydrogen Integrated Greenhouse Horticultural (HIGH) Energy Project

Funded by Ontario’s Independent Electricity System Operator (IESO) through the Hydrogen Innovation Fund, the HIGH Energy project explores using wind-powered green hydrogen to decarbonize the greenhouse sector in Southwestern Ontario. Led by the University of Windsor in collaboration with Kruger Energy and the Ontario Greenhouse Vegetable Growers, this research assesses the technical and economic feasibility of blending hydrogen with natural gas to provide clean heat and power for large-scale horticultural facilities.

The Part 1 report identifies a lowest levelized cost of hydrogen (LCOH) of CAD $4.6/kg using a 10% H₂/90% natural gas blend at $33.07/MWh electricity rate. Economic viability is optimized at 5 days of storage and 2,190 annual operating hours, with analysis covering 9 scenarios across various blending ratios and delivery methods.

Key Techniques

Research Scope

• • 9 scenarios across H₂ blending ratios (10%, 20%, 100%) and delivery methods
• • Wind-powered PEM electrolysis coupled with 200MW wind farm
• • Levelized cost, payback period, IRR, and discounted payback analysis
• • Market analysis and regulatory review for Ontario grid integration

Project Links

The Alternative Deicer Paradox

Demonstrates that "green" alternative road deicers (calcium chloride, magnesium chloride, potassium chloride) mobilize trace metals (Cu, Pb, Zn) at higher ecotoxicity than conventional sodium chloride, contradicting the prevailing assumption that alternatives are environmentally safer. Submitted to the Journal of Environmental Management. An interdisciplinary collaboration across Western University and ÉTS Montréal (Université du Québec).

Key Techniques

Key Findings

• • Alternative deicers mobilize Cu, Pb, Zn at higher ecotoxicity than NaCl
• • Challenges the "green alternative" narrative in municipal de-icing policy
• • Data-driven evidence for policy recommendations on deicer selection
• • Submitted to the Journal of Environmental Management (Elsevier), 2026

Microplastics Detection via ML

DETR object detection output: microplastics classified as Glitter (1.00) and Fragment (0.99) from Great Lakes samples

Using DETR (Detection Transformer) and CNN-based deep learning to automate the detection and classification of microplastics gathered from the Great Lakes. Applying state-of-the-art object detection models for rapid, accurate identification of microplastic particles in environmental samples.

Key Techniques

Methodology

• • DETR (Detection Transformer) for microplastic detection
• • CNN-based classification of microplastic types
• • Great Lakes environmental sample analysis
• • Object detection pipeline for automated identification

Medical Imaging AI

Brain image generation using Denoising Diffusion Probabilistic Models (DDPM), created by Dr. Petrov

14+ years of medical imaging research at the Robarts Research Institute. Expertise spans deep learning for MRI motion correction, metal artifact reduction in CT, cardiac micro-CT imaging, and diffusion-based generative models for brain image synthesis.

Key Techniques

Selected Work

• • Deep learning motion correction for multichannel MRI (NMR in Biomedicine, 2024)
• • Brain image generation with DDPM diffusion models
• • Metal artifact reduction in CT scans
• • Cardiac micro-CT iterative reconstruction
• • 3D cardiac imaging with portable cone-beam scanner

The Cost of Retreat from an Inherited Hardline Posture: A Decision-Analytic Diagnostic for Multi-Issue Bargaining

Belief evolution trajectories: DP-optimal vs bluff-and-retreat bargaining strategies

A decision-analytic benchmark for multi-issue bargaining where a single focal actor chooses bundles (tariffs, side payments, security concessions) against a reduced-form opponent with a two-type acceptance rule. The key mechanism is credibility depletion: visible retreat from hardline positions reduces future bargaining leverage, whereas consistent conduct rebuilds it. Joint work with Matt Davison.

Key Techniques

Key Results

• • Priced cost of a climb-down from an inherited hardline record: 0.08 to 0.18 in utility units under high stakes
• • With a moderate record, the forward-looking policy coincides exactly with the myopic one, zero value for foresight where none is needed (a validity check)
• • Against a fully strategic responder the sender-side effect nearly vanishes, a limit characterised rather than concealed
• • Submitted

Belief-State Dynamic Programming for Typed Submarine Payload Management under Regime-Switching Threat Uncertainty

Formulating the reserve-management problem for submarine-launched unmanned underwater vehicles (UUVs) as a partially observable stochastic control problem. The mothership carries a typed inventory of ISR, mine countermeasures (MCM), and anti-submarine warfare (ASW) vehicles, with a shared handling bottleneck preventing simultaneous launches. The hidden threat follows a three-state continuous-time Markov chain (peacetime, tension, conflict). Joint work with Matt Davison.

Key Techniques

Key Results

• • Optimal policy yields mean reward of 12.54 vs 8.36 (myopic) over 4,000 Monte Carlo missions
• • Hold region occupies ~77% of belief simplex at mission start
• • ISR dominates in tension; MCM/ASW emerge only under high conflict posterior
• • Identifies a cooldown parity effect: terminal alternation in the optimal policy boundary driven by the handling cycle, not regime dynamics
• • Submitted to the IEEE Journal of Oceanic Engineering

Optimal Stockpiling of Defence-Critical Minerals under Regime-Switching Price Dynamics

Regime-switching price dynamics for nickel, copper, and cobalt (1990–2024)

Canada produces ten of the twelve minerals NATO classifies as defence-critical yet lacks an operational framework for timing procurement under a hard readiness requirement. This work formulates procurement as a constrained control problem, minimizing discounted cost subject to a hard no-shortage requirement, using a two-state hidden Markov model of monthly log-returns and dynamic programming over the joint price–inventory–belief state. Calibrated to World Bank nickel/copper and Investing.com cobalt data. Joint work with Matt Davison; submitted to INFOR: Information Systems and Operational Research.

Key Techniques

Key Results

• • Constrained policy makes zero shortage a property of the design, at 9.4% (Ni), 9.1% (Cu), 7.0% (Co) below the cheapest same-information zero-shortage heuristic
• • The constrained solver converges within 0.1% and costs 24.8% less than a penalized alternative
• • Cobalt's thin futures series fails two pre-deployment screens, a condition diagnosed before deployment rather than assumed
• • Submitted to INFOR: Information Systems and Operational Research

Health-Aware Reserve Management for Submarine-Launched UUVs: A Dual-Belief Model and the Sufficiency of Point-Estimate Control

Extends a single-belief UUV threat model with a second uncertainty dimension, the internal health state of stored vehicles, and asks whether tracking the full health-belief distribution is worth its cost. The central finding is that it is not, in this regime: a certainty-equivalent controller acting on the most-likely health tier captures essentially the entire value of the full dual-belief policy. Joint work with Matt Davison.

Key Techniques

Architecture & Key Results

• • External: three-state threat regime (peacetime / tension / conflict) tracked with a Wonham filter
• • Internal: latent health tiers from a physics-informed digital-twin library with Weibull reliability
• • A certainty-equivalent controller on the most-likely tier matches the full dual-belief policy exactly under common random numbers (sixfold sensor-noise sweep)
• • A failure-aware four-tier policy beats the failure-blind three-tier policy by up to +2.78 return units on the degraded stratum; under review

Common Noise-Erosion Laws for Refractory Detectors

Refractory detectors in biology and optics share a vulnerability: after each event the device is briefly blind, leaving a hole of exact zeros in the event-train autocorrelation that noise erodes. A small set of common laws governs this erosion across detector classes, forming a design principle: from record length, event rate, and refractory scale alone, one can predict the tolerable noise before acquiring data. Joint work with Matt Davison.

Key Techniques

Key Results

• • An identifiability limit: the binarized hole sees error rate and event density only through their product, so a detector's native noise cannot be read from its own hole; calibrated counts recover both
• • A mechanism law: erosion collapses onto a parameter-free master curve when the within-hole hazard is flat, departing in a signed direction otherwise
• • Validated on 20 real cardiac, photonic, and neural trains spanning 2.8×10⁶ in timescale (pooled R² = 0.985 under controlled corruption)
• • Converts a target fidelity into a pre-acquisition error budget; submitted to Nature Communications

Mortgage Renewal Negotiation Under Information Asymmetry, Search Frictions, and Credible Outside Options

A theoretical framework for Canadian mortgage renewal as a repeated bilateral renegotiation between a household and its current lender, translating the credibility-depletion mechanism of the companion bargaining paper into mortgage observables. Joint work with Matt Davison.

Key Techniques

Key Results

• • Seven testable hypotheses linking preparation, timing, credibility, and lender response latency to renewal outcomes
• • Calibrated to a negotiable margin of roughly 95 to 125 basis points below posted rates
• • A terminal-region result: close enough to the maturity date the negotiation channel is effectively closed regardless of preparation, verified across a 150-configuration sweep
• • Under review

A Transparent Calibration-Based Agent for Human Bargaining

A minimal, closed-form, transparent agent for the ultimatum game, offered as an alternative to opaque language-model agents: a smooth acceptance rule plus a small set of additive threshold shifts, each calibrated from one human single-factor study. Because published human acceptance curves disagree widely, it calibrates to effect sizes rather than absolute levels. Joint work with Matt Davison; funded through the NFRF Exploration program.

Key Techniques

Key Results

• • Published human acceptance curves disagree by up to 35 percentage points at a given offer, motivating calibration to effect sizes rather than absolute acceptance
• • Two informative failures reported honestly: single factors do not combine additively, and the agent cannot reproduce the effect of experience
• • A cross-game threshold-mechanism check in a repeated market with asymmetric information, without refitting
• • Under review

Terminal Parity: Discrete Impulse Control with Refraction

Identifies and characterizes a structural signature in finite-horizon optimal control problems: when an agent faces a discrete handling cooldown (refraction constraint), the optimal policy boundary oscillates near the terminal horizon. This "terminal parity" effect is driven by the interaction between the countdown clock and the handling cycle, not by the underlying regime dynamics. A theoretical refinement of the cooldown-parity phenomenon first identified in the TKMS UUV payload management paper. Joint work with Matt Davison.

Key Techniques

Key Results

• • First documented treatment of terminal-horizon oscillation in impulse-control literature
• • Oscillation traced to handling cycle, independent of regime dynamics
• • Abstract structural result applicable beyond naval operations
• • Under review