Mana Games

[ChrisM47]

Hey everyone,

Long-time Tennis Elbow player here. I've been lurking on the tournament threads for a while, and I have a question that's been on my mind.

A friend of mine is into machine learning and has been experimenting with a tennis prediction algorithm on the side. Nothing serious, just a hobby project where he pulls match data, engineers features per player (recent form, surface adjusted stats, rolling serve metrics, head-to-head) and trains a small model that spits out a win probability for upcoming matches. We compare his model picks against ours every week during the tour.

The overlap with what Tennis Elbow simulates under the hood is what I find fascinating. You've got the same problem, predicting point-by-point outcomes from a bunch of player attributes. Obviously the game designers have their own ways of modelling it, but I'm curious how close a purely data-driven approach would get.

Has anyone here played around with predictive models outside the game? Or has anyone read about how Tennis Elbow itself handles the probability model under the hood? I'd love pointers either way.

[Jira]

Well we're planning on adding prediction stuff in the future in our website https://tenniselbowhub.live/ in the future with Marco(Ambience). For the time being we're still implementing stuff little by little. Today we added a tournament predictions tab but is for the users to share their prediction in upcoming official tour matches. Maybe later if we introduce some other features we could do that

[manutoo]

ChrisM47,
It seems you're talking about the offline World Tour, but posted in the Online Tour forum.
TE4 doesn't really predict CPU vs CPU match results, it simulates them. In other words, the goal isn't to estimate an outcome from external data, but to generate the match from the game's internal player attributes and match logic.
Also, unlike a pure prediction model, this simulation is part of the World Tour gameplay, so the user can influence it through in-match decisions.

[ChrisM47]

Reading back the thread now and realising I never properly answered Jira and manutoo. Will do that in a follow up post in a moment, just need to gather a couple of numbers first.

In the meantime, quick technical update for context: I ended up going much simpler than I had planned. First a logistic regression over surface, recent form, head to head and a couple of fatigue proxies as a baseline. Then LightGBM on the same features for a small but real bump. Service hold rate weighted by surface is doing most of the heavy lifting honestly.

[ChrisM47]

Apologies for the delay on the follow up, I went away from the project for a few weeks and only came back to it recently.

@Jira, that is a really nice update on tenniselbowhub.live, I had not seen the prediction tab go live. The user-driven angle is honestly the smarter move because the moment a tab is filled with model picks, half the readers stop reading and the other half complain about the ones that miss. Curious whether you plan to weigh user picks by past accuracy or just keep them as raw counts.

@manutoo, you are completely right that I posted in the wrong section, and the simulation versus prediction distinction is really the heart of the question. What my friend's algo is trying to do is the opposite direction of what TE4 does internally, instead of generating a match from player attributes, we infer outcome probabilities from past results, surface, fatigue and a few rolling features. So they are kind of cousin problems with the data flowing the other way around.

The cleanest signal so far is still service hold rate weighted by surface, which beats most of the fancy features I tried to bolt on. We've been comparing weekly outputs against bookmaker closing lines just as a sanity check, and against the official tour rankings drift, with a small write up of our weekly notes here when we have time to format it. Most of it ends up sitting in a Google Doc anyway.

If anyone has tried encoding intra-match momentum or tilt without overfitting on tiny samples, would love to compare. That is the part where Tennis Elbow players probably have better intuition than my training data does.