What Actually Predicts a Rugby Injury (And What Doesn't)

Two seasons, one squad, and a model that catches 4 in 10 injuries in its top risk bucket.

Last season a senior tight forward at Kintetsu got hurt in week 29. By the time the medical staff confirmed it, the model I’d been building had already flagged him as the highest-risk player in the squad two weeks earlier. Risk score: 83%. The next-highest-risk player that week didn’t get injured.

That’s one example. I’m not claiming the model is clairvoyant. I’m claiming that across two seasons of daily GPS, weekly medical reports, and 200-odd injury events, the patterns that show up before injuries are real and they’re not the ones most people in rugby spend their time on.

This post is the first of a small series on what those patterns are, in plain English. Later posts will cover how I’d build this into a Tuesday-morning dashboard, and what changes when you can see two years of daily GPS instead of one season of weekly rollups. This one is the foundations: what predicts an injury, what doesn’t, and what you’d start measuring differently if you took the data seriously.

What the data is

Two seasons of Kintetsu Liners data. 78 players, 14,400 days of GPS at the per-player-per-day level, 402 distinct injury episodes parsed out of two years of weekly medical reports plus a structured injury log from the prior season. Each injury is tagged with body region (hamstring, calf, knee, ankle, shoulder, head/face, low back), mechanism where I could read it (strain, sprain, contusion, fracture, concussion), severity grade where the staff recorded it, and side.

The model takes a player on a given day and asks: what’s the probability they get hurt in the next 7 days? I trained it on weekly rolling windows so it can’t peek at the future. Every result here is out-of-sample.

The headline

If I rank every player-day in the squad’s last season from highest predicted risk to lowest, and look at the top 18% of those rankings, I capture 43% of the injuries that happened in the next 7 days.

A coach reading that should hear: roughly 1 in 5 days you’d be paying attention to about 4 in 10 of the injuries before they happened. That’s not a science-fiction crystal ball, but it’s the difference between blanket caution across the whole squad and targeted load adjustment on the players who actually need it.

The model isn’t right every time. The top 18% bucket has an injury rate of about 8% (that’s roughly 1 in 12 player-days in that bucket actually result in an injury within a week). The bottom 80%, by contrast, has an injury rate around 2%. So a player flagged in the top bucket is roughly 3-4 times more likely to get hurt in the next week than a player in the bottom one. That’s the lift.

What’s actually predictive

Here’s the part that surprised me the least and the most. The single strongest signal in the model is how recently the player was last hurt. If they were in rehab last week, they’re not at the top of the list this week. If they had a niggle three months ago that flared into a soft-tissue strain, they’re more likely to break again in the next month than someone with a clean history.

That sounds obvious. Coaches know recurrence is real. But the size of the effect surprised me: among the 80-odd features I gave the model, the recency-of-prior-injury feature had nearly twice the importance of the next strongest predictor.

The next strongest predictors:

1. Total training time over the last 28 days. High cumulative volume, especially when chronic load has been climbing for weeks.
2. Career injury count from the prior season. Players with more injuries last year are more likely to have one this year, even controlling for current load.
3. Recent high-speed running ratio. A player whose recent HSR is high relative to their own 4-week average.
4. Recent high-intensity acceleration ratio. Same idea but for accel efforts above 2.5 m/s².
5. Days since the last hard session. This one was a surprise. The model flags players who haven’t done anything hard recently, especially right before they get hurt. That fits the load-management paradox: the player getting deloaded because the staff already saw warning signs.

What’s notably not in the top 10:

• Bronco PB. The aerobic test had almost no standalone signal in this dataset.
• Squat and bench PBs. Strength tests, on their own, don’t predict injury.
• BMI, weight. Body composition has roughly nothing to say at the daily level.
• Distance covered today. Single-day load is too noisy. Rolling 4-week chronic is what carries the signal.

Different injuries have different signatures

This was the finding I didn’t expect. If you take the three weeks before each injury and look at what the player’s training looked like, the shape of the pre-injury window depends on what they’re about to get injured.

I’ve put hamstring, ankle, and knee side by side because they tell three different stories.

Hamstring strains are preceded by low load. The player isn’t running hard, isn’t doing much volume, and their acute-to-chronic ratio is around 0.80, well below the “normal” baseline of around 1.0. That’s almost certainly because by the time they strain, they’re already being managed. They’ve been pulled back by the staff for a niggle. Hamstring injuries in this data are mostly recurrence-driven, not acute-overload-driven.

Ankle injuries are preceded by high HSR. Players who go on to roll an ankle are running about 50% more high-speed metres in the lead-up than the squad average. That fits the mechanism: ankle sprains happen during fast changes of direction, in tackles taken at speed, and when the foot lands awkwardly during a sprint.

Knee issues happen at sub-maximum velocity. Players who go on to a knee injury are running at about 80% of their personal max velocity in the weeks before, compared to 89% for healthy weeks. They’re not flying. They’re in the middle range where contact, twisting, and changes of direction live. That makes sense too: most knee injuries in rugby come out of contact, not from top-speed running.

Calf and low-back injuries look more like the average. Neither high-load nor low-load. They’re driven by chronic factors that don’t show up cleanly in 3-week pre-windows.

So a region-blind injury model loses signal. A model that knows what kind of injury you’re trying to predict can make sharper calls. The hamstring-specific model in this dataset hits about 60% accuracy on its top-ranked predictions, even with only a dozen events. That’s higher than the all-injury model on the same days.

Contact injuries can’t be predicted from load

This is worth saying clearly because the load-monitoring industry sometimes implies the opposite. Concussions, fractures, dislocations, contusions: the contact category. Those show no meaningful signal from training load at all in this dataset. The model trained to predict shoulder injuries (mostly dislocations and AC sprains) hovered around random.

The corollary is that the injuries you can actually predict from load are the soft-tissue, non-contact ones: hamstring strains, calf strains, tightness, tendinopathies. That’s also where the literature consistently finds load-injury signal.

If you tell me 60% of your squad’s injury list is broken bones and concussions, the load model is going to have a hard time helping you. If 60% of your list is soft-tissue, this kind of model is the right tool.

The 4-week ACWR doesn’t pull its weight

The standard load-management metric in rugby is the acute-to-chronic workload ratio: this week’s training load divided by the average of the last four weeks’. Coaches and sport scientists have been tracking it since about 2016. Spike above 1.5, get injured. That’s the version of the story most people know.

In this dataset, the 4-week version of that ratio is mediocre. On its own it predicts injuries with about 57.5% accuracy, barely above chance. But if I extend the chronic window from 4 weeks to 12 weeks, the same ratio becomes meaningfully better:

That’s a six-point jump in predictive accuracy from doing nothing more than widening the window. The intuition: a player’s “normal” load isn’t really set by the last four weeks. It’s set by the last three months. Especially across a season with bye weeks, finals taper, and pre-season ramp, four weeks isn’t long enough to be a stable baseline.

So if you’re tracking ACWR, track the 12-week version. If you only track one chronic window, track 12. The 4-week version isn’t worthless, but it’s the inferior choice.

Preseason peak does something specific

A lot of S&C coaches operate on the principle that you build the player up in preseason so they can survive the season. The data says that’s directionally right but only on one specific axis.

I looked at every preseason peak metric I could engineer: biggest weekly distance, biggest weekly HSR, peak velocity hit, days at ≥85% of own max velocity, days running at ≥85% of game speed. Most of them have basically no relationship with the player’s in-season injury rate. But one has an interesting one:

Each preseason day where a player ran above 85% game speed translated to about 110m of additional high-speed running per match in-season. That’s a strong signal in a small sample (R² of 0.26 on 14 player-seasons), and it points at the right thing: the players who do high-intensity work in preseason carry that capacity into matches.

So the preseason hypothesis “load them up to make them durable” is mostly wrong if you measure durability as injury rate. But the version “load them up to expand their match capacity” is right. Players who hit harder preseasons do more in games. They don’t necessarily get injured less.

Peak weekly acceleration efforts in preseason were the one mild protective factor: more preseason accels, slightly fewer in-season injuries (R² of 0.11). The other peak metrics didn’t move the needle.

What I’d actually do on a Tuesday

If I were running this on a real squad next season, here’s the dashboard I’d build:

1. Per-player risk score, daily. Not a green/amber/red traffic light. A continuous risk percentile. A coach can read “you’re in the top 20% of risk this week” and act on it. They can’t act on a colour.
2. Top three drivers per player, in plain language. “Days since last hard session: 1 (model thinks you’re under-recovered). 4-week training time: 3rd-highest in squad. Last injury: 5 weeks ago.” A coach can look at that and decide whether to intervene.
3. Region flag if the player’s pattern matches a region’s fingerprint. If a back’s recent HSR is unusually high and they have an ankle history, that’s worth a conversation. If a forward’s load has dropped and their ACWR is around 0.8 with a hamstring history, that’s a different conversation.
4. 12-week chronic load on every player, plotted across the season. Not the 4-week version. Not just this week’s spike. The slow-moving 12-week curve is the actual baseline.
5. No load model for contact injuries. I’d treat tackler-load and contact density as a separate problem with different inputs (mechanism, technique, opponent). The injury model is for soft-tissue.

The honest part

A few things I want to be straight about.

This is one squad, two seasons. Around 200 events. That’s not a lot when you spread them across body regions. The hamstring-specific finding is built on 12 events. The knee-specific result on 7. I’d want every one of those numbers to reproduce on a second team’s data before I’d bet meaningful resources on them.

The model gets to about 70% accuracy at predicting any injury in the next 7 days. That’s useful but not magic. Most of the rugby injury models in the literature land at 65 to 75% accuracy when they have wellness questionnaire data: daily soreness scores, sleep, mood, RPE. I didn’t have those for this dataset. The models that hit the upper end of that range almost always have wellness inputs.

And the load-management bias is real. Some of what looks like “low load before injury” is the staff already responding to early warning signs the player gave them. The data we observe is post-intervention. A perfect model would need to know what the staff did and didn’t do, which I don’t have records of.

So: the findings are stable in this dataset, the model works, the patterns make biological sense, but the science isn’t done. The next step is the same model running on a second team’s data, ideally with wellness collected daily, ideally over three or four seasons.

What changes from this

The findings I’m willing to put weight on, in order of confidence:

1. Recurrence is the strongest predictor of future injury. A player with a recent injury history is the highest-leverage population for monitoring and rehab investment.
2. 12-week chronic windows beat 4-week ACWR. This is a free lift in any team currently tracking the standard ratio.
3. Different body regions need different monitoring. A region-blind dashboard is leaving signal on the table. Hamstring, ankle, and knee fingerprints are distinguishable.
4. Contact injuries don’t respond to load monitoring. Stop pretending they do. Spend that energy on tackle technique and contact load instead.
5. Preseason intensity translates to in-season match capacity, not durability. That changes how you justify preseason work to a head coach.

If you’re an S&C coach reading this, my Tuesday-morning suggestion is: pull your historical injury data, mark which were soft-tissue and which were contact, and ask how many of the soft-tissue ones followed a player who’d been managed for something in the previous month. If it’s most of them, your prevention strategy probably starts with rehab quality more than with screening tools.

The next post in this series is what the model would have actually said for the squad’s biggest injury moments last season, week by week. That one’s a worked case, not a numbers post.