Models, markets, and the interfaces between them.

Single-Factor Modeling

Cross-sectional factor research dashboard for ranking, diagnostics, buckets, risk, and backtest review.

• Configurable metric, lookback, and bucket controls propagate through every panel.
• Signal quality combines IC, Spearman, Pearson, slope, R², monotonicity, and spread.
• Backtests report Sharpe, Sortino, drawdown, Calmar, win rate, and terminal wealth.

AutoML Workbench

A no-code platform to train, rank, explain and deploy many scikit-learn, XGBoost and LightGBM models on a point-in-time equity panel or a user's own tables, scored with purged walk-forward validation so future data never leaks in.

• Led the path from a ported research library to production: purged walk-forward CV with embargo and a multi-date holdout, SHAP drivers, blended ensembles, a Redis/RQ worker queue, and an audit of 438 data limits so that every cap became a named refusal.

Notebook Dashboards

One click turns a Python research notebook into a shareable, interactive dashboard: typed inputs, scheduled reruns, live embeds, and server-side pivots and cross-filtering over every row of the data.

• Built an in-house DuckDB pivot engine that streams answers over Arrow. The worst-case query once compiled for 28 s and peaked at 10.8 GiB; every limit is now measured and refused by name, never silently truncated.

Operator Catalog

Eight separate analytics and ML operator registries unified into one generated, hash-sealed catalog. Every operator is a typed parameter form that resolves to runnable notebook code, plus private operators generated from any Python library in the user's own kernel.

• Deterministic, schema-contracted projection (22 JSON Schemas) with CI conformance runs, content-addressed fail-closed delivery, and one resolver kept identical across Python, TypeScript and kernel implementations.

Point-in-Time Market Data Platform

QuantLink AI's survivorship-free market-data foundation: vendor feeds in ClickHouse under Dagster, a universe that keeps delisted companies, ~180M rows of calculations proven against their sources, and a replayable dbt lakehouse.

• A daily incremental lands ~5.9M rows in ~14 minutes, every calculation table ships only after matching an independent oracle, and the lakehouse keeps exact provider bytes for replay.

SQL-Native Backtester

A backtesting engine written entirely as dbt SQL models on ClickHouse: no Python loop, look-ahead blocked by construction, every dollar of P&L reconciled to its trades, and rebuilt in production every week.

• Look-ahead blocked structurally and by a fuzzed compile-time gate; a per-symbol P&L identity test, a hand-traced trade (84/84 daily rows, zero difference) and a toy fixture covering every exit path pin correctness.

Notebook AI Agent

A grounded AI agent that writes and edits research notebook cells against a live Python kernel, and checks its own work before any edit applies: code parsed, SQL validated and grounded to real datasets, numbers matched to cited evidence.

• Built the runtime underneath too: isolated Jupyter kernels with signed, renewable grants, live recovery with no cell replay, and a digest-pinned production executor verified end to end.

QuantLink MCP Platform

An MCP layer where AI agents research, evaluate, and refine strategies with a quant desk's tools, plus a governed data server giving agents and notebook kernels fail-closed, entitlement-checked access to 789 datasets.

• Large results stay out of agent context: Parquet snapshots served by handle, DuckDB SQL follow-ups with no second provider call, verified end to end on staging.

Options Workspace

A full options analytics workspace on QuantLink's stock page: 104 views spanning a mirrored chain, an interactive 3D implied-volatility surface, skew and term structure, built and shipped across the data layer, API and web app.

• Profiled the volatility-surface pipeline stage by stage: a C-accelerated loader for a 6,594-field catalog cut fresh-process hydration from 6.2 s to 1.4 s, and failed streams return a structured, traceable error.

Agentic Alpha Engine

End-to-end alpha engine for US public equities, built entirely on public data and designed to be driven by any MCP-capable AI agent.

• Spec-driven YAML datasets feed dbt staging and marts, with generated tests and a JSON-LD catalog.
• French, AQR, and JKP factor libraries; PyPortfolioOpt and skfolio optimization; tested end-to-end in CI.

American Options & Securitized Products

Numerical methods for complex derivatives and fixed income: American options, barrier options, MBS, and G2++ short-rate models.

• Binomial and trinomial lattices with dynamic programming for path-dependent pricing.
• Monte Carlo with antithetic and control variates for variance reduction.

QQQ Regime Strategy

Long/short QQQ strategy gated by trend, volatility, and tail-risk regimes before any exposure is taken.

• Separates market direction from risk state before taking exposure.
• Compact audit trail from rule logic to tearsheet.

ETF & Crypto ML Strategy Research

Sliding-window LightGBM research for Bitcoin, Ether, and QQQ with oscillators, liquidity, and macro-trend features.

• EWMA, RSI, Amihud, lottery, bid-ask spread, and SPY trend features.
• Beat the benchmark in live paper trading after transaction costs.

FOMC Sentiment Toolkit

FinBERT-FOMC research turned into an installable package and CLI covering transcript extraction, LLM-assisted labeling, fine-tuning, embeddings, novelty scores, and market regressions.

• Hawk/dove lexicon baseline, per-speaker profiles, and walk-forward LightGBM evaluation against SPY.
• Interactive dashboard for factor series, a per-paragraph meeting explorer, and speaker analytics.

AI Exposure Factor Suite

ETF-based market-implied AI return factors: an AI Theme factor and an AI Infrastructure factor, built point-in-time for firm-level AI exposure estimation.

• F_AITheme (AI ETFs ex-semiconductors) and F_AIInfra (semiconductor / AI infrastructure) daily factors.
• QC diagnostics against Fama-French factors; rolling betas for firm-level exposure work.

CEO Tracker

Self-hosted app tracking CEO speeches and press releases from top US firms, with company mentions via spaCy NER and FinBERT sentiment.

• Postgres + Celery pipeline for ingestion and scoring.
• Executive language turned into trackable company signals.

Early Warning System

A market-crash probability model for US equities: regime and tail-risk signals distilled into a single live probability.

RL Statistical Arbitrage

Stat-arb combining copula dependence, theory-implied correlation, graph structure, and reinforcement learning.

Stat Arb Optimization & Risk Analytics

Copula-based US equity pairs strategy with RMT-denoised and shrinkage covariance estimation, transaction-cost-aware convex portfolio optimization, and granular performance analytics.

Implied Volatility Strategies

Equity strategies built on implied-volatility data from the options market, with volatility-surface signals treated as tradable edges.

LEAPS as Stock Substitutes

Empirical long/short equity backtest using LEAPS in place of the underlying stock: escrowed-spot Black-Scholes, a calibrated IV proxy, and a daily-accounting engine with exercise, expiry, and ruin mechanics, on public data only.

• Staged pipeline: download → pricing validation gates → single-instrument lanes → long/short portfolios → sensitivity grid → REPORT.md.
• Engine tests cover NAV reconciliation, call-exercise and put-parity reopens, expiry settlement, split-scaled frictions, and dollar-neutrality.

Factor Investing with Options

Factor portfolios constructed with deep in-the-money options instead of stocks, with leverage and downside asymmetry as design inputs rather than afterthoughts.

Modified Momentum

Replication of the classic momentum portfolio, then adapted for current market structure.

MAG7 Concentration Effects

IAQF 2025 research on whether SMB factor behavior decoupled after concentration shifted into mega-cap winners.

• Post-COVID concentration vs. historical size-factor assumptions.
• Framed as both a portfolio risk and a research-design problem.

LightGBM Technicals

Technical signals built and combined with LightGBM for single equities and ETFs.

WorldQuant Multi-Region Alphas

Intraday and medium-frequency alpha work across US, China, Europe, and UK markets via WorldQuant BRAIN.

• Fundamental, analyst, short interest, news, options, and price-volume features.
• Signals tuned to different market structures and liquidity conditions.

Loss-Making Companies Factor Model

A risk-neutralized factor model for loss-making companies: a LightGBM and Elastic Net ensemble for the regime where traditional valuation signals break down.

Complex Itô Random Walk

A 13-phase research program: an exact-law framework for stochastic correlation with closed-form transition densities, barrier laws, and spread/basket pricing.

• Act I bridges complex polar form, Brownian quadratic variation, and Itô's lemma to random walks.
• Act II prices spread/basket options, occupation-clock derivatives, and de-peg derivatives in closed form.

Options Analytics Dashboard v3

Rebuilt options analytics: a 6-leg strategy builder with American exercise, five American pricing models side by side, a Greeks lab, and exotics via Monte Carlo.

• Early-exercise premiums, convergence charts, and American implied volatility back-solved from live quotes.
• Dash backend with a Mantine/React frontend; pricing library vendored from the American_Options project.

Oil Geopolitical Risk Index

A composite 0–100 geopolitical risk index for crude oil, distilled from prices, prediction markets, news flow, positioning, and live vessel traffic at chokepoints.

• Fail-soft async collectors for CFTC COT, Baker Hughes rigs, EIA inventories, and macro rates.
• FastAPI backend with a terminal-finance ECharts frontend and 60-second auto-refresh.

Kalshi Market Data & Mentions Pricing

Data pipeline and pricing engine for Kalshi's "Mentions" markets: whether a word will be said during a speech or an earnings call, and how many times. A hurdle model prices the contracts; a FastAPI + Mantine dashboard explores them.

• Shrunk Beta-Binomial reference-class base rates × truncated negative-binomial repetition, calibrated so E[N|N>0] matches history.
• Optional LLM and topic-similarity adjusters; mid-event live pricing anchored to the hurdle prior.

Polymarket Earnings Markets

Public-data pipeline for Polymarket earnings-beat markets: discovery, backfill, a DuckDB warehouse, and a T-2 probability signal backtested as an event-driven equity strategy.

WTI Multi-Agent Forecasting

Five independent LLM-agent runs forecasting the maximum daily EIA WTI Cushing spot through 2026, with a reproducible Quarto paper on forecast-revision behavior after peer exposure.

MAPO: Multi-Asset Optimizer

One optimization API across equities, crypto, futures, FX, and treasuries. PyPortfolioOpt, skfolio, and Riskfolio-Lib wrapped behind a unified interface with native solvers, plus a no-lookahead walk-forward backtester and a CLI.

• Exact max-Sharpe as a homogeneous SOCP under box, asset-class, and turnover constraints; auto-fallback to min-vol when excess returns are non-positive.
• Risk models from EWMA and Ledoit-Wolf to PCA factor and nearest-PSD repair; asset-class caps (e.g., crypto ≤ 10%) are first-class constraints.

report-forge

Quarto-powered publishing engine, driven by agents over MCP: Markdown and local chart PNGs in, typography-grade PDF, HTML, and DOCX out, with mechanical quality gates that refuse to ship bad exhibits.

• Editorial cover infographics (verdict band, key-point cards, scenario strip) compacted onto page one; native Exhibit numbering and spreadsheet-style tables in PDF and HTML.
• Production renderer for QuantFlow outlooks (MSFT, GOOGL, META, NVDA); headless notebooks with persistent kernels in v0.2.0.

Spark-MCP

Creativity and institutional-learning MCP server for quant research: every trial logged (failures included), significance deflated for the full trial count, and a MAP-Elites map of unexplored research space. A peer engine computes; Spark judges.

• Exact-once trial counting: retries never inflate significance; validated trajectories distill into replay-gated skills with a human veto.

kizoNLP

Financial NLP library: sentiment engines (DeBERTa-v3 financial, zero-shot NLI, SetFit few-shot), a modern embedding registry (Qwen3, gte-modernbert, BGE, E5), and map-reduce summarizers for long documents, all exposed as an MCP server.

Quant Dashboard Systems

Dashboards for backtesting, screening, alpha review, financial models, and macro risk monitoring, built at Wolfe Research and elsewhere, with every recent dashboard launched from a custom-built Mission Control hub.

IMC Prosperity 2025

Algorithmic trading submission for IMC Prosperity 2025: market-making and positioning under live-style constraints.

Quant Toolkit Contributions

Forked and extended quant research tooling: portfolio analytics with quantstats, plus stochastic-process simulation for research reproducibility.