The paper asks how to recover a stable, interpretable, state-dependent stochastic discount factor when returns and conditioning information are high-dimensional. The starting point is a regularized conditional quantile portfolio program. It estimates one portfolio together with a conditional quantile surface, using elastic-net penalties to stabilize both the portfolio weights and the state-dependent quantile coefficients.

The central distinction in the talk is between three objects. The quantile portfolio is a return-side object. The raw expected-shortfall envelope uses residual ranks to identify tail states, but it is not automatically a pricing kernel. Only after projecting that tail-focused object onto the no-arbitrage pricing set do we obtain an admissible stochastic discount factor.

Empirically, we apply the workflow to monthly Fama–French portfolios and U.S. macro-financial states. The main message is diagnostic rather than structural: macro-state patterns, projection failures, and boundary cases reveal where tail emphasis and pricing feasibility align, and where the data put stress on the candidate pricing kernel.

My presentation was titled From Training Machines to Teaching Students to Think with AI. The central argument was simple: methods education is about intellectual autonomy. In statistics, data analysis, and data science, students do not need manual routines as an end in themselves. But they do need enough understanding to ask sharper questions, evaluate AI-generated outputs, expose assumptions, and remain accountable for analytical work.

AI lowers operational friction. It helps with code generation, debugging, method explanations, documentation, and workflow design. That is useful. At the same time, some of the friction it removes was also where students learned to reason. The teaching challenge is therefore not to ban assistance, but to keep students inside the reasoning loop: ask, try, explain, and govern.

I discussed this from the perspective of large introductory statistics and data-analysis courses with about 800 students per year. At that scale, AI-era teaching quickly becomes an infrastructure question. Good learning design needs repeated practice, visible feedback, randomised and varied tasks, transfer across contexts, and privacy-aware ways for teaching teams to see where students struggle.

That is why my practical conclusion was institutional rather than merely individual. Universities need more than access to AI tools. They need deployment infrastructure, AI inference capacity, orchestration skills, and governance routines that let lecturers build and maintain course-specific support responsibly. Agentic coding can make small teaching apps easier to build, but hosting, authentication, monitoring, data access, privacy, and maintenance remain institutional responsibilities.

The presentation also connected naturally to the broader governance discussion in the expert group. Teaching and governance are two sides of the same problem: students need autonomy, and institutions need capacity. Adoption without capacity risks hidden dependence; responsible use requires that students, lecturers, and universities know where judgement and accountability remain.

My main takeaway from the talk was this: teach students to think with AI, not to watch AI think for them.

The current version is described in a non-peer-reviewed working paper. The model is a state-dependent goal-arrival generator. Goals are treated as jumps in a continuous-time score state; the transition law at full time gives a joint distribution over scorelines. Win-draw-loss probabilities, exact-score decisions, tournament simulations, and total-goals forecasts are then functionals or decisions derived from that distribution.

Update, August 3, 2026: the working-paper PDF was revised after review feedback. The revision sharpens the evaluation against generator, machine-learning, and market-implied benchmarks; the public forecast below remains the June 8 freeze candidate.

The forecast below is a freeze candidate dated June 8, 2026. It uses historical international matches available up to June 7, a football-ranking.com strength snapshot from June 4, and no player or lineup covariates. The total-goals side forecast excludes shootout goals and in-match penalty goals.

Headline Predictions

Method In Brief

The generator estimates goal-arrival intensities as functions of team strength, current score state, and match timing. It then propagates those intensities into score probabilities. For each fixture, the forecast stores both result-class probabilities and exact-score probabilities. The revised paper compares the event-time generator with final-score likelihoods, calibrated machine-learning benchmarks, and market-implied probabilities under fixed information sets.

This matters because the forecast target changes the evaluation. A model can be good at win-draw-loss probabilities and less good at exact scores; betting-market prices can dominate close-to-kickoff probabilities while being a different information set than a strict pre-tournament forecast. The working paper therefore treats the forecast object, information set, and scoring rule as part of the method.