Categories on Philipp D. Dubach | Quantitative Finance & AI Strategy

AI

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Coverage spans AI economics, enterprise deployment strategy, and the measurable effects of AI on labor markets and industries.

A recurring theme is the gap between AI benchmarks and real-world utility. Posts on model evaluations and scaling law economics examine whether trillion-dollar infrastructure bets translate into proportional capability gains. The enterprise AI coordination problem argues that most organizational failures stem from integration, not model quality, while the enterprise agent stack maps how orchestration, memory, and tool-use layers are separating into distinct infrastructure tiers.

On AI economics, the OpenAI standalone P&L analysis examines unit economics at scale, and the recommendation algorithm economics behind Netflix and Spotify shows why compute-optimal recommendations diverge from business-optimal ones. The a16z vs AQR AI valuation debate tests whether current AI infrastructure spending can generate proportional returns.

The AI and labor displacement analysis tracks measurable job market shifts, while The Impossible Backhand examines where AI quality ceilings make human domain expertise more valuable, not less. Applied AI work includes sentiment-based trading signals using news embeddings and hands-on agent-building projects.

Economics

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Economics on this site is empirical and quantitative: agent-based simulations, behavioral experiments, game theory applied to actual decisions. Less debate about what models should look like, more results from running them.

The flagship piece, agent-based modeling of wealth distribution, implements the Affine Wealth Model from Bruce Boghosian’s group at Tufts. Random transactions between equal agents produce Pareto-distributed outcomes, and the simulation reproduces 27 years of U.S. wealth data within 0.16% average error. About 10% of agents go negative, matching the share of U.S. households with negative net worth. The post then runs counterfactuals on what a wealth tax of 1% to 5% does to the distribution: the result is a steady-state equilibrium where the wealthiest agents hold at most 3 to 4 times their initial allocation.

Behavioral coverage starts from Kahneman, Thaler, and Ariely but lands in concrete policy. The zero price effect in transit shows why making public transport free generates demand shifts much larger than the equivalent fare reduction would predict, with implications for fare-free transit experiments now running in Tallinn, Luxembourg, and Kansas City. The economics of mobile money traces the M-Pesa playbook across African and Asian markets, looking at when network effects dominate regulatory friction.

Game-theoretic analysis appears where signaling and auction dynamics dominate outcomes. The Super Bowl ad piece breaks down the per-spot bidding equilibrium between brand-budget incumbents and growth-stage challengers, and asks why broadcast advertising’s ROI math still works for some categories and not others.

The thread across these posts is a preference for models that produce testable predictions over models that produce explanations.

Investing

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Investing here means valuation, not prediction. The frame is Mauboussin’s: identify situations where market price and intrinsic value diverge measurably, then size positions to the conviction that the gap will close. Not “where is the S&P going next year.” More “is this asset worth what it costs, and on what assumptions.”

Valuation coverage includes Everything is a DCF Model, applying Mauboussin’s framework across asset classes, and the a16z vs AQR debate over whether AI’s trillion-dollar infrastructure bets will generate proportional returns. The SaaSpocalypse paradox examines how the market simultaneously prices AI capex failure and software disruption.

Market structure posts cover passive investing’s structural fragility, prediction market adverse selection, and the Kalshi debate over where speculation ends and investment begins.

Historical case studies include Michael Burry’s newsletter for dot-com parallels, Buffett’s retirement as a case study in capital allocation, and personal 2026 portfolio allocation as a worked example of thesis-driven positioning.

Macro

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Analysis of monetary systems, central bank policy, and geopolitical realignment, focused on structural shifts rather than cyclical noise.

The two-part Pozsar’s Bretton Woods III series examines how commodity-backed reserve currency alternatives challenge dollar hegemony, tested against three years of subsequent data. Related macro coverage includes the $12.6 trillion repo market and its opacity, Japan’s $5 trillion foreign asset position and yen carry trade unwind risk, and Fed dual mandate tensions under conflicting inflation and employment signals.

The Stewart-Thaler debate dissects why the economy “feels bad” despite good headline data, connecting cumulative price levels to consumer sentiment. Europe’s payment infrastructure breakup tracks the structural shift from card network dominance toward sovereign account-to-account rails.

Geopolitical coverage includes the rise of middle power realism reshaping trade blocs and Britain’s post-Brexit strategic limbo between US and EU spheres.

Medicine

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Analysis of drug development economics and medical research, with particular depth in GLP-1 receptor agonists and the competitive dynamics between Novo Nordisk and Eli Lilly.

The GLP-1 coverage tracks the class from pharmacology to market impact: oral peptide delivery challenges explains why semaglutide’s 1% oral bioavailability shapes the competitive landscape, while Novo Nordisk’s post-patent strategy analyzes the amycretin pipeline against the semaglutide patent cliff. The Ozempic and fast food analysis examines downstream demand effects on consumer industries. Emerging research on GLP-1 for addiction treatment extends the receptor’s relevance beyond metabolic disease.

On the pharma economics side, Isomorphic Labs and AI drug discovery examines whether AI-designed drug candidates can beat the Phase II wall, and the Novo Nordisk crash analysis dissects what the CagriSema trial failure and 72% drawdown mean for the GLP-1 market.

Where medicine intersects with data science, posts apply quantitative methods to health data, including continuous glucose monitoring analysis and glycemic response prediction models.

Quantitative Finance

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Data-driven analysis of financial risk, volatility, and the mathematics underlying market behavior.

Central to this category is the variance tax, the half-sigma-squared drag on compound returns that explains why volatility matters more than most investors realize. The companion piece on the long volatility premium synthesizes evidence from One River, Goldman Sachs, AQR, and Universa on whether tail hedging is a compensated factor or an expensive insurance policy.

On market structure, It Just Ain’t So challenges the Gaussian assumption underlying standard financial models, examining fat tails and their implications for risk management. The private equity beta analysis tests whether the illiquidity premium survives after adjusting for leverage and mark-to-market smoothing.

Applied work includes tensor-based market modeling as an alternative to Monte Carlo simulation, and the mathematics of provably fair casino games applying probability theory to verifiable randomness.

Tech

me@philippdubach.com (Philipp D. Dubach) — Mon, 01 Jan 0001 00:00:00 +0000

Tech here means hands-on engineering: building working systems, breaking them, writing about what shipped and what didn’t. Not opinion pieces about the industry, not vendor reviews. Code that runs, with the tradeoffs documented.

Project work covers computer vision (a card counter trained on OpenCV and an audit of F3ED, the NeurIPS 2024 tennis-shot detector), applied machine learning (visualizing PyTorch gradients, modeling postprandial glucose response with XGBoost), and health-data engineering (a Python library for the Dexcom CGM).

The site itself is a working case study. The blog runs on Hugo, deploys to GitHub Pages on push, sits behind Cloudflare. Four separate Cloudflare Workers handle the parts a static site can’t: a security-headers Worker emits CSP and HSTS plus an Accept-aware content-negotiation layer that returns index.md to crawlers sending Accept: text/markdown; two AI Workers running Llama 4 Scout 17B auto-post to Bluesky and Twitter every 15 minutes, deduplicating via KV; a GoatCounter proxy serves the “most read” data in the footer.

Posts in this section show the engineering decisions, not the cleaned-up summaries. When a deploy broke, what broke. When a model underperformed, by how much.