Systematic strategy

What I do

I design investment-strategy research and test whether its results survive: whether a backtest holds when the parameters move, and when the strategy runs forward on data it has not seen. Below, to help you judge whether I fit your project: my selected work in this field, the methods and tools I would bring to it, my advising and outreach, and my review and teaching. If I do, the Connect page says how to start.

Selected work

Academic journals, chapters, and policy writing

Peer-reviewed articles

  • "Financial Regulations and Price Inconsistencies across Bitcoin Markets," with Sofia Vivanco. Information Economics and Policy.

    Every bitcoin is identical, so a lasting price gap between markets must come from the markets themselves. Across 11 exchanges, together 26% of global bitcoin trade volume, we find that exchanges that do not require customer identification to open an account are more likely to deviate from representative market prices, so know-your-customer rules can have a non-negligible effect on the bitcoin market.

    Data. Daily bitcoin prices in US dollars from 11 exchanges, together 26% of global bitcoin trade volume, and a bitcoin price index, June 2014 to July 2015, with 22 characteristics of each exchange collected by hand from its own pages, such as whether it requires identification to fund an account.

    Methods. Unit-root tests: DF-GLS, KPSS. Cointegration tests of the law of one price: Johansen trace, Pesaran-Shin-Smith bounds. Toda-Yamamoto Granger causality. OLS with robust standard errors and date fixed effects, regressing each exchange's price deviation on its characteristics.

    Read the paper

Working papers/work in progress

  • "Bitcoin Reveals Exchange Rate Manipulation and Detects Capital Controls"

    The data used to detect currency manipulation and capital controls are low-frequency, costly, lagged, and possibly mismeasured. I show that bitcoin prices, once adjusted for bitcoin-market-specific effects, approximate unofficial exchange rates at a daily frequency, and so reveal both whether a distortion exists and how large it is, including short interventions that would otherwise go undetected. The result also confirms that bitcoin is used to get around capital controls.

    Data. Daily transaction-weighted bitcoin prices for 17 currencies on four exchanges, June 2014 to September 2015, with official exchange rates.

    Methods. Bitcoin-implied exchange rates, adjusted for bitcoin-market trends using the euro as the benchmark. Cointegration tests of each bitcoin rate against the official rate, Johansen trace and Pesaran-Shin-Smith bounds, to detect capital controls. Toda-Yamamoto Granger causality. Validation against the Chinn-Ito capital-openness index and the IMF's exchange-rate regime classification.

    Read the paper · Media coverage: The Telegraph, Bitcoin News, Finance Magnates, Breaker Mag

  • "Manufactured Premia: How Universe Construction Creates the Crypto Cross-Section," with Chelsea Canon.

    Crypto factor studies often build their universe from the tokens still alive at the end of the sample. We rebuild four published factor families on a point-in-time universe, booking a total loss when a token dies, and run identical code on a twin universe built the published way. The size and volatility premia exist only on the twin; momentum survives honest data but not a 50 basis-point trading cost.

    Data. Daily prices, market capitalizations, and trading volumes for crypto tokens, January 2020 to June 2026; each month's 100 largest tokens, using only information available that month; for momentum, every token above $1 million in market capitalization, also chosen point in time.

    Methods. Factor portfolios on two universes with identical code; Newey-West t-statistics; trading-cost analysis.

    Work in progress

  • "Robust Search: Exploiting Tolerance To Optimize Trading Strategies"

    Choosing a trading strategy's parameters by backtest rewards whichever setting fits the past. I present a parameter sweep that searches only among robust settings: settings that keep passing when every parameter moves at once within a declared tolerance. A lattice sized by the tolerance places a point in every region wide enough to hold a tolerance box, with the fewest points that can, at a cost known before anything runs. A local search then certifies points inside those regions. A capped cascade searches the certified ground for the best performer, and a final phase names the highest-ranked candidate that passes every declared test. On a synthetic surface with two robust regions, a decoy and a spike, the first two phases certify both robust regions in every run. At six parameters, none of twelve alternatives given the same budget and judged by the same test does so in any run.

    Data. A synthetic surface with a known answer, in two or six parameters, each on the unit interval, with four square regions that score equally well: two good regions, a decoy that fails a second test at every point inside it, and a region too narrow to hold a tolerance box. Each evaluation returns a score and a second value for the decoy test.

    Methods. A lattice with points twice the tolerance apart covers every region that holds a tolerance box; a neighborhood test then certifies points in the cores of the regions found; a cascade of shrinking low-discrepancy placements, with a cap on the points each round keeps, searches the certified ground; and a final phase names the highest-ranked candidate that passes every declared test. Compared with twelve alternatives: a Sobol scatter, four global optimizers, an optimizer of the average score over the corners of the tolerance box, a practitioner's heuristic, and five robust black-box optimizers, over eight seeds.

    Work in progress

Talks and presentations

Central banks and international institutions

  • Federal Reserve Bank of Dallas: "Bitcoin Reveals..."
  • De Nederlandsche Bank: "Bitcoin Reveals..."

Academic conferences

  • Royal Economic Society: "Bitcoin Reveals..."
  • Western Economic Association: "Financial Regulations..."
  • Southern Economic Association: "Bitcoin Reveals..."
  • Eastern Economic Association: "Bitcoin Reveals..."
  • Midwest Macroeconomics Meeting: "Bitcoin Reveals..."
  • Money, Macro and Finance Conference: "Bitcoin Reveals..."
  • International Trade and Finance Association: "Bitcoin Reveals..."
  • Society for Economic Measurement: "Bitcoin Reveals..."
  • Liberal Arts Macroeconomics Workshop: "Bitcoin Reveals..."

University seminars

  • Lafayette College: "Bitcoin Reveals..."
  • Hobart and William Smith Colleges: "Bitcoin Reveals..."

Policy and industry forums

  • Monetarium 2, Washington: co-presented the asset-backed currency design.

Reports, projects, and applications

Strategy and portfolios

  • Asset-backed currency: designed, simulated, and tested a modular, CPI-indexed asset-backed currency on AWS compute; sole designer, with a selectable basket of backing assets. Developed its inflation-tracking portfolio model, which draws on roughly 4,167 assets, with beta screens, comparisons of loss functions, rolling validation, deduplication by principal components, and selection over millions of candidate portfolios.
  • Product proposals: converted early investment concepts into documented, testable product proposals and supported evidence-based decisions to launch, revise, or retire them.
  • Strategy-family study: scoped and ran a study of 37 base strategies (moving-average strategies and signal strategies, such as RSI mean reversion and breakouts) and 1,052 variants, with fees and taxes, and a go/no-go assessment for a bitcoin trend-following product.
  • Bitcoin-gold rotation: directed a synthetic tokenized-gold price history, a reconciliation of differences between backtests, and a walk-forward study of bitcoin-gold rotation, which found no selection skill.
  • Attention-driven rotation: directed backtests driven by page views and volume spikes, point-in-time, with dead coins kept.
  • Risk signals: researched a stablecoin-flow risk signal and directed its strategy harness, write-up, and deployment specification.
  • Red-teaming proposed strategies: matched the robustness tests to the kind of strategy: strategies with tunable parameters, which can be overfit mechanically, were tested for parameter robustness; definitional strategies, such as a revenue-based one, were tested by ablation, removing tokens, roles, and signals and varying basket size, weighting, rebalance timing, and execution delays. Made the tests a reusable, rerunnable template.

Infrastructure improvements

  • The Robust Search method: developed it into a 69-page methods document, a working package, a standalone parameter-selection engine benchmarked against one live product and three proposed strategies, and team training. Used in: Robust Search.
  • Computational speed-ups: directed speed-ups to the inflation-tracking model, deduplicating generated portfolios and updating the selection and generation code. The selection step fell from six hours to five minutes, an estimated full run on an AWS cluster from 3.1 years to 4.8 days, and the portfolios generated rose from about 25,000 in a three-day run to about 300,000 in one day.
  • Shared backtester: contributed a robustness module and a cost model to the team's shared backtester, and designed a standardized PDF strategy report.
  • Parameter-sweep pipeline: designed an automated nine-step pipeline that searched the parameter settings of 42 strategies (30 moving-average strategy variants and 12 signal strategies, such as RSI mean reversion, breakouts, and Bollinger squeezes), narrowing the parameter search in six rounds, testing whether the best result beat a benchmark or was luck (Hansen's SPA test), and writing the reports.
  • Scenario-stress engine: directed its build: a hidden semi-Markov regime generator with Monte Carlo ensembles.

AI deliverables

  • Mining the literature with AI: built an AI-assisted research pipeline across more than 6,000 academic papers, turning an initial 600 studies into testable investment ideas.
  • LLM companion packages: built them for three proposed products.

Student projects

  • Designing student data projects on crypto market data: forecasting the bitcoin price when sources report different daily prices; the most profit, in hindsight, from arbitrage and price moves across three exchanges, net of fees; a trading strategy written as code to run on an exchange; and one token's returns, volatility, momentum, and investor attention, following Liu and Tsyvinski. Used in: Cryptoasset Markets.

Industry research

  • Lead Research Economist, Confusion Capital, a capital management company in the Reserve ecosystem: quantitative research developing on-chain index products.

Methods and tools

Models and computational simulation

  • Parameter search (which settings are worth testing?): Sobol and quasi-random search, adaptive grids that narrow the search in rounds, and multi-objective optimization. Sobol search used in: Robust Search.
  • Scenario simulation (how would it fare in conditions not in the data?): a hidden semi-Markov regime generator with Monte Carlo ensembles.
  • Portfolio simulation (what is left after costs?): with explicit fees, taxes, and benchmark comparison.

Data and measurement

Data construction and validation

  • Reconciling prices across sources: daily bitcoin prices from 11 exchanges, compared with a reference price to measure each exchange's deviation. Used in: Financial Regulations....
  • Collecting what no database holds: 22 characteristics of each exchange, such as its customer-identification rules, gathered by hand from the exchanges' own pages (used in Financial Regulations...), and one day's trading volume for every pairing, on every exchange, of the 50 largest cryptocurrencies.
  • Validating a new measure: bitcoin-implied exchange rates in 17 currencies, checked against the Chinn-Ito capital-openness index and the IMF's exchange-rate regime classification. Used in: Bitcoin Reveals....
  • Building point-in-time universes: ranking tokens by market capitalization each month using only information available that month, booking a loss of 100% when a token dies. Used in: Manufactured Premia.

Econometrics

Quantitative investment methods

  • Systematic strategy design (what rules should a strategy follow?). Used in: Robust Search; Manufactured Premia.
  • Backtesting (how would it have performed?). Used in: Robust Search; Manufactured Premia.
  • Robustness tests and local-stability checks (can a backtest result be trusted?): Hansen's SPA test, probability of backtest overfitting (PBO), deflated Sharpe ratio, White's reality check, stationary block bootstrap, walk-forward and holdout validation, Morris sensitivity screens, look-ahead causality checks, and pre-test gates, with parameter advisers and champion-challenger registries. Local-stability checks used in: Robust Search.
  • Machine-checked proofs in Lean 4 (do the guarantees hold?): directed 13 proofs for the guarantees of a parameter-search method and 3 for the tests behind a stablecoin-flow risk signal.

Computational research tools

  • Validating AI-assisted code: research code moved to LLM-assisted development, with the rewrite validated by reproducing the prior baseline portfolio before features were added.
  • Multi-agent research system: designed and direct a multi-agent AI research system for replication, data pipelines, and writing.
  • Comparing models: directed comparisons of Claude models and multi-model review cascades, scored for accuracy, unsupported claims, and cost, and used separation between agents as a quality control.
  • Auditable AI workflows: built workflows for research, simulation, writing, and code review, and connected research findings to product positioning, controlled launches, and customer feedback.
  • Testing with synthetic users: directed synthetic-user personas to test a natural-language strategy workflow (identifying potential strategies from user prompts, and which strategies may appeal); the workflow had little success at either.
  • Auditing code with AI: directed a multi-agent audit of the team's code and testing results.
  • Avoiding lock-in: set up a consolidated, Git-based local workspace to avoid lock-in to a single IDE or model provider.
  • AI research workflows: LLM research pipelines; multi-agent orchestration with Git-based handoffs; source-grounded LLM assistants; Claude Code and Codex.
  • Programming, computing, and documents: AWS compute; MATLAB and Stata; Git; LaTeX; Python and other languages through LLM-directed development.

Advising and outreach

Guest lectures

  • Undergraduate: Fintech (University of Illinois at Urbana-Champaign).

In the media

Review and teaching

Evaluation and review

Review

  • Reviewed investment methodologies for model risk, survivorship bias, transaction costs, liquidity, and implementation feasibility, translating findings into portfolio and deployment recommendations.

Refereeing

  • Served on 3 conference selection committees for the Crypto Asset Lab Conference.

Courses taught

Except where marked co-taught, I designed and taught these courses, from determining the curriculum to creating the homework and exam problems, and grading them or overseeing their grading. For more details see my academic website.

  • Master's: Cryptoasset Markets, Financial Mathematics, University of Chicago.

    A course for students with no prior knowledge of crypto. It covers the technology, the markets, and the players, and tests where the analysis tools of traditional finance work, or fail, in cryptoasset markets.

    Topics. Blockchain technology and history; centralized markets and the reliability of reported exchange data; spatial and triangular arbitrage; decentralized finance, automated market makers, and impermanent loss; crypto returns, factor exposure, and momentum; and the roles tokens play in projects. Guest speakers from industry.

    Assessment. Five data projects, most done in groups, including a trading strategy written as code to run on an exchange. Students post their code and data to a shared folder, so each learns from the others' work.

    Syllabus