Sports Analytics Weekly by kubeia.io - 17/2026

At the ICE 2026 gaming conference in Barcelona, our investigators posed as operators of a startup sportsbook. We told SRAD’s sales team — repeatedly and explicitly — that our target markets were Vietnam, Thailand, Indonesia, and China. Every one of these countries bans online gambling. Not one SRAD salesperson told us no. Instead, an Asia-focused sales executive walked us through product offerings tailored to each illegal market, bragged that SRAD “serves everyone,” and offered to introduce us to the Yabo Group — China’s largest illegal gambling operator, whose Cambodian call centers are staffed by trafficked and enslaved workers. He warned us that Yabo’s people didn’t attend ICE because “they would be hunted down.” Then he offered to make the introduction anyway. This is who SRAD is.

The set piece revolution remains the story of the season in the Premier League and shows no signs of slowing down. Goals from set pieces are still elevated. Corner kicks and long throws continue to account for more or less the entirety of this effect. Since the last Expecting Goals newsletter pinpointed these two tactics as the core of the new set piece vision, more discussions and analyses have focused on these situations.

Expected Goals (xG) models have become a central tool in football analytics, widely adopted across broadcasting, coaching, and performance analysis. The paper highlights that xG is now “a standard part of match facts” and increasingly used by coaches such as Mikel Arteta and Arne Slot to evaluate performance. Among providers, Hudl-StatsBomb positions its model as the “most accurate xG model”, which motivates the authors to critically examine this claim.

Football analytics has become increasingly sophisticated at measuring shots. We can estimate chance quality, compare finishing skill, model goalkeeper positioning, and understand which teams consistently outperform their xG totals. All of that has genuine value.

But goals are rarely created at the instant a player strikes the ball.

They are usually built in the seconds beforehand: a forward dragging a centre-back out of position, a midfielder receiving between lines for one touch too long, a full-back arriving unnoticed on the blind side, or a defensive line shifting half a second later than it needed to.

The finish is visible. The construction is often hidden.

For decades, tactical analysis in football was built on observation. Coaches, scouts, and analysts watched the game, interpreted shapes, and described what they believed they saw. One team looked compact. Another controlled midfield. A striker occupied defenders well. A back line was too deep. Much of it was insightful, but much of it also lived in language rather than evidence.

Then tracking data changed the landscape.

Once every player could be located multiple times per second, football gained something it had never fully possessed before: a measurable map of collective behaviour. Suddenly, spacing could be quantified. Synchronisation could be tested. Defensive reactions could be timed. Tactical analysis no longer had to rely purely on description.

The authors published it at a moment when the game was beginning to move from intuition-led tactical commentary toward data-supported tactical understanding. Long before tracking departments became standard, before pressing metrics entered mainstream discourse, and before clubs openly discussed machine learning, the authors asked a question that still sits at the centre of modern football analytics:

What if tactics could be measured through movement patterns rather than explained after the fact?

It sounds simple now. At the time, it was a statement about where the game was heading.

Human pose estimation has become a cornerstone of modern computer vision, powering applications from fitness tracking apps and sports analytics to gesture-based interfaces and medical rehabilitation. At its core, keypoint estimation is the task of detecting specific anatomical landmarks on the human body, the nose, shoulders, elbows, wrists, hips, knees, and ankles, and connecting them into a skeleton that represents the person’s pose.

In September 2025, Ultralytics announced YOLO26, the next-generation YOLO model optimized for edge computing, robotics, and mobile AI. Among its specialized task heads, YOLO26-pose brings several architectural innovations to keypoint estimation: Residual Log-Likelihood Estimation (RLE) for more accurate keypoint localization, end-to-end NMS-free inference for simpler deployment, and the MuSGD optimizer for more stable training dynamics. In this guide, we walk through the theory, architecture, benchmarks, and a hands-on implementation of YOLO26 keypoint estimation on images and videos.

Tabular foundation models such as TabPFN and TabICL don’t need to be trained to perform regression or classification. What they do is called in-context learning. What used to be the training data now becomes the context data at prediction time.

This post explores the idea of context data and contrasts it with “classic” training data. Does moving from training data to context data change how we model? Does it enable something new?

Let’s dive in.

In Episode 100 of the SBC Podcast, I welcomed a special guest to mark the milestone, in the form of Antonino (Ant) De Rosa, a former Pinnacle trader and now operator of a large scale professional betting group.

This episode goes deep into the reality of sharp betting. Not theory, not models in isolation, but how markets actually move, how sportsbooks react, and why execution is often the biggest edge in the modern game.

Ant shares his unique journey from elite level Magic: The Gathering player to being recruited by Pinnacle, where he specialised in live NBA trading. He explains how his edge was never about knowing sport better than others, but about predicting behaviour, understanding where the next bet would come from, and how the market would react.

The conversation then shifts into the mechanics of running a serious betting operation today.

From scaling across multiple sports, managing hundreds of accounts, and handling real world challenges like liquidity, restrictions, and non payment. Ant is clear that having a strong model or idea is not enough, if you cannot get money down, the edge has no value.

This is a rare, honest look at the sharp end of betting in 2026, where success comes from combining insight, discipline, and the ability to execute at scale.

In Episode 101 of the SBC Podcast, I’m joined by Phil Smith, a full time football bettor specialising in antepost (futures) markets and exchange trading.Phil shares his journey from early losses and trial and error, through to building a profitable approach combining Betfair trading and more recently, antepost betting.We go deep into how antepost betting works in practice, where the real edge comes from, and why volume, structure and discipline are key to long term success.The conversation also explores the growing challenges facing bettors today, including Phil’s own experiences, from being permanently removed from Betfair after years of activity, to being locked out of funds and bets for months during an affordability check with Sky Bet.This is a practical, honest discussion about modern betting, where theory often takes a back seat to execution, access and adaptability.