Sports Analytics Weekly by kubeia.io - 11/2026

This study investigates how physical contexts differ across Europe’s top five leagues and how athletic ability is leveraged differently to generate player output. The authors develop models designed to project player performance when transferring between leagues based on physical metrics, age, team quality, and position. The central premise is that league environments impose distinct physical demands, and that these contextual differences materially influence performance translation.

The paper starts from a familiar limitation in football analytics: most defensive value remains invisible because many important defensive actions happen without direct contact with the ball. A defender may never register a tackle, interception, or clearance, yet still prevent danger simply by occupying the right space and discouraging a pass. The authors frame this around cover shadows, the defensive occupation of passing lanes that forces opponents away from valuable options. They argue that traditional event statistics fail precisely because they only record realized actions, while a successful cover shadow often removes an action before it exists.

For years, football formations were treated like labels.

4-4-2.

4-3-3.

4-2-3-1.

A quick graphic before kickoff, a TV overlay, a line in the match report.

But formations are rarely just shapes. They are tactical choices shaped by context, squad quality, physical demands, and even by what coaches believe the modern game now requires.

And if you zoom out long enough, you are able to see interesting things. Not one formation replacing another, but an entire league slowly changing how it distributes risk, control, and attacking presence.

That is what makes this LaLiga study particularly interesting.

Rather than focusing on one team, one coach, or one season, it looked across 3,420 matches and 6,840 starting formations over nine seasons, asking a simple question:

How has elite football actually changed structurally over time?

And the answer is more nuanced than the usual “everyone moved from 4-2-3-1 to 4-3-3.”

Nice looking backtests are cheap now.

This is worth sitting with for a moment.

In the age of AI, a beautiful backtest proves almost nothing.

The probability that some parameter combination produces an amazing equity curve approaches certainty as the number of combinations you try increases.

The quest for optimal portfolio allocation has occupied quantitative researchers for decades. Markowitz gave us mean-variance optimization in 1952,¹ and since then we’ve seen Black-Litterman, risk parity, hierarchical risk parity, and countless variations. Yet the fundamental challenge remains: markets are dynamic, regimes shift, and static optimization methods struggle to adapt.

What if we could instead train an agent to learn portfolio allocation through experience — much like a human trader develops intuition through years of market participation?

Enter reinforcement learning (RL). Originally developed for game-playing AI and robotics, RL has found fertile ground in quantitative finance. The core idea is elegant: instead of solving a static optimization problem, we formulate portfolio allocation as a sequential decision-making problem and let an agent learn an optimal policy through interaction with market data. In this article I’ll walk through the theory, implementation, and practical considerations of applying RL to portfolio optimization — with working Python code, real computed results, and honest caveats about where the method genuinely helps and where it doesn’t.

Tabular foundation models (TFMs) are a paradigm shift from traditional tabular ML: They are transformer-based architectures pre-trained on synthetic data. There is no classic training step. Instead, TFMs predict the test data in a single forward pass of combined training and test data without any parameter updates (in-context learning).

These last few weeks of deep-dive have reshaped how I think about TFMs and tabular ML as a whole. I won’t claim I can predict the future. I’ve been completely wrong before, like about how good AI would become at coding. Instead of predictions, here are a few scenarios of increasing impact of TFMs (levels) on everyday tabular ML work.

Let’s dive in.

Foundation models for tabular data are powerful.

They are also dangerously easy to misjudge.

And in structured domains — especially high-stakes ones like finance, insurance, healthcare — misjudgment isn’t harmless.

It leads to fragile systems.

It leads to misplaced confidence.

It leads to real consequences.