Systematic Equity/ Quantitative Research/ Applied ML
Associate Portfolio Manager at Parametric Portfolio Associates (Morgan Stanley). I work on systematic equity SMAs — portfolio construction, rebalancing, and the Python analytics our team uses to run them day to day.
About
An optimizer result only becomes a portfolio once it clears tax lots, client restrictions, corporate actions and a transaction cost budget. Closing that gap is most of my job, and it is why I spend about as much time writing code as running analysis.
Building and rebalancing custom core SMAs against client-specific benchmarks, balancing tracking error, tax impact and transaction costs against the constraints in each mandate. Scenario analysis on flows, rebalances and benchmark changes, written up so advisors can act on it.
Network analysis and machine learning applied to market structure. My undergraduate honors thesis forecast bull/bear regimes from the topology of industry correlation networks. The habit stuck: look for the structure first, then test carefully whether it predicts anything.
Python applications used in daily PM workflows, plus the data-integrity checks behind 400+ benchmarks. Selected as an AI Ambassador at Parametric, where I help colleagues adopt AI tools and build their own.
Experience
I started where risk numbers are produced, learned what breaks them, and moved to the desk that has to act on them. Each step has taken me closer to the investment decision itself — from measuring exposure, to building the portfolios that carry it, toward the quantitative research and systematic trading I want to work on next.
Parametric Portfolio Associates · Morgan Stanley Investment Management
Construct and rebalance $9Bn+ of systematic equity custom core SMAs against client-specific benchmarks, trading tracking error against tax and transaction costs. Build the PM-facing Python analytics and the data-integrity layer under 400+ benchmarks.
Morgan Stanley
Daily Rates and FX sensitivity analysis: risk-engine output aggregated under a delta–gamma approximation to estimate scenario PnL across market environments. Built the Python and SQL pipelines that monitored those sensitivities against limits.
Morgan Stanley
Portfolio analytics and reporting across a $200Bn+ lending portfolio in SQL, Python, R and VBA, including the error-checking routines behind senior management reporting.
Research
Work on market structure, machine learning and portfolio construction. The honors thesis below asked whether the topology of an industry correlation network carries a forecastable signal about the market regime that follows it — at short horizons it appears to, over longer ones the signal largely disappears. New work is added here as it is finished.
Honors Thesis · UNC Chapel Hill Economics · May 2020
Historical U.S. equity returns are partitioned into bull and bear markets using a top-down turning-point algorithm. For each regime a correlation network is built over the 49 Fama–French industry portfolios, reduced to a minimum spanning tree, and summarised by its topological features. Those features — degree distribution, in-component degree, Kruskal stress and density, alongside the moments of S&P 500 returns — are fed to a bank of machine learning classifiers under 10-fold cross-validation. Across 16 combinations of estimation window and forecast lag, a Fine Tree on 1-month windows forecasting one week ahead reaches 91.7% accuracy; over longer horizons a Kernel Naive Bayes on 3-month windows at a 12-month lag holds 66.7%.
Method
STEP 01
Partition the regimes
Hanna's top-down turning-point algorithm over daily S&P 500 prices, 1978–2019: 30 bull, 26 bear, 5 rallies, 7 corrections.
STEP 02
Correlate the industries
Within each regime window, the Pearson correlation matrix over daily log returns of the 49 Fama–French industry portfolios.
STEP 03
Build the tree
Correlations become Euclidean distances, dij = √(2(1−Cij)), reduced to a minimum spanning tree scored by Kruskal stress-1.
STEP 04
Extract topology
Degree distribution, in-component degree, stress and density per network — plus the standard deviation, kurtosis and range of market returns.
STEP 05
Classify & validate
Trees, KNN, SVM, ensembles and Naive Bayes across 16 window×lag combinations, each under 10-fold cross-validation.
Forecast accuracy by estimation window and lag
Sixteen model configurations, best classifier per cell. The signal is concentrated: only the shortest horizon breaks meaningfully away from chance.
| Window | Lag | Accuracy | Best model |
|---|---|---|---|
| 1 month | 1 week | 91.7% | Fine Tree |
| 1 month | 2 weeks | 77.1% | Cosine KNN |
| 1 month | 3 weeks | 62.5% | Subspace Discriminant |
| 1 month | 4 weeks | 62.5% | Naive Bayes |
| 3 months | 1 month | 62.5% | Fine Tree |
| 3 months | 3 months | 52.1% | Boosted Tree |
| 3 months | 6 months | 58.3% | Linear Discriminant |
| 3 months | 12 months | 66.7% | Kernel Naive Bayes |
| 6 months | 1 month | 64.2% | Linear Discriminant |
| 6 months | 3 months | 54.7% | Cubic SVM |
| 6 months | 6 months | 58.3% | Fine Tree |
| 6 months | 12 months | 52.1% | Coarse KNN |
| 12 months | 1 month | 54.7% | Coarse Gaussian SVM |
| 12 months | 3 months | 52.8% | Coarse KNN |
| 12 months | 6 months | 54.2% | RUSBoosted Tree |
| 12 months | 12 months | 52.1% | Coarse KNN |
Accuracy under 10-fold cross-validation. Rallies and corrections were excluded from training — too few observations for a valid sample — so the task is a bull/bear binary.
Where the winning model gets it right
Fine Tree, 1-month window, 1-week lag. Four misclassifications out of 48 regimes, split evenly between the two error types.
| Actual | Predicted bear | Predicted bull |
|---|---|---|
| Bear | 21 | 2 |
| Bull | 2 | 23 |
For context, Kole & Dijk (2010) forecast the same one-week-ahead bull/bear call with Markovian logit models at 89.3%. This model reaches 91.7% with an AUC of 0.92 — on 48 regime observations, so the difference sits well inside the noise.
Honest limitations. The MST minimisation criterion did not fully converge, so the trees may not be the most efficient representation of each network; and several algorithm-partitioned regimes are only days long — artefacts of black-swan events sitting too close to the following rally to be separately forecastable. Both are named in the paper as the first things to fix. With 48 regimes in the sample, the accuracy figures carry wide error bars either way.
Recurring trade and portfolio-analysis scenarios — security flows, rebalances, benchmark reconstitutions, tax-loss activity — summarised with an LLM copilot and clustered into a categorised template library. The work is in the categorisation: which scenarios recur often enough, and are distinct enough from one another, to earn their own template. Portfolio-manager correspondence to financial advisors then starts from a scenario-matched draft carrying the relevant risk and tax diagnostics rather than from a blank page.
Co-authored at the Morgan Stanley firmwide GenAI Hackathon: generate text embeddings from 10-K filings, surface the fundamentals that matter to a credit view, and automate the first draft of the credit memo. Drafting time fell by about 70% in the hackathon evaluation.
Teaching Assistant for data science at Duke's Fuqua School of Business and the Department of Statistical Science. Research Assistant at Duke's Center for Advanced Hindsight, a behavioural economics lab.
Building
PM-facing Python applications I have built at Parametric. Small, specific tools that run inside a live investment process and are maintained in shared repositories.
Rules-based extraction and validation of trading scenarios from structured advisor notes — regex parsing with error handling, so an ambiguous instruction is flagged rather than silently misread.
Benchmark reconstitution impact analysis — adds and drops, resulting turnover, and which SMAs are affected — so rebalance work can be prioritised by impact rather than by list order.
The shipped form of the scenario template library: portfolio-manager correspondence to financial advisors, drafted from copilot-summarised, categorised templates matched to the recurring trade and portfolio-analysis scenarios behind it.
Consistency checks across securities and screening groups using hash-table comparison, catching restriction mismatches before they reach a portfolio that shouldn't hold the name.
A distributive trade status reporting system with automated scheduling and validation checks, streamlining the communication loop between portfolio managers and traders on what has actually filled.
Dashboards tracking ingestion across 400+ benchmarks, reconciling corporate actions against Bloomberg, FactSet and Refinitiv. Not glamorous, but every rebalance downstream depends on it.
Toolkit
Education
/ MAY 2022
Duke University
Teaching Assistant for data science courses at the Fuqua School of Business and the Department of Statistical Science. Research Assistant at the Center for Advanced Hindsight.
/ MAY 2020
University of North Carolina at Chapel Hill
Second major in Psychology, minor in Mathematics. Graduated with Highest Distinction and the Herbert Brown Mayo Financial Economics Research Fellowship.
Off the desk