Training Load and Injury Risk

Your plan was written by something that has never felt your shin. Runna doesn’t know that your left tibia ached for four days in March. Garmin Coach sees your HRV and your Training Readiness score but not the fact that you moved house last weekend and slept on a sofa. A ChatGPT-built plan doesn’t even know what you did before week one.

So the question that matters isn’t “is this a good plan.” It’s narrower and more answerable: is this plan asking my body to absorb more change per week than it has recently demonstrated it can absorb? That question has arithmetic behind it, and the arithmetic is the acute:chronic workload ratio. It also has a large amount of overconfident nonsense attached to it, which is worth clearing out first, because a number you trust wrongly is more dangerous than no number at all.

What the acute:chronic workload ratio actually measures

The calculation is trivial. Take your load over the last 7 days (acute). Take your average 7-day load over the last 28 days (chronic). Divide.

ACWR = acute 7-day load / mean weekly load over 28 days

Two versions exist and they give different answers. Coupled ACWR includes the current week inside the 28-day chronic window, so this week’s load appears in both numerator and denominator. Uncoupled ACWR uses only the three preceding weeks for the denominator. Coupling drags the ratio toward 1.0 and systematically hides spikes. Most consumer apps use coupled without telling you.

The famous thresholds (0.8 to 1.3 “sweet spot”, above 1.5 “danger zone”) come from Tim Gabbett’s work in rugby league, cricket and Australian rules football, popularised in his 2016 “training-injury prevention paradox” paper. Those are collision sports with squad-level GPS data, weekly fixtures and a single load metric that everyone shares. Not a 41-year-old running four times a week around a park.

The evidence in running is thinner than your app implies

This part gets skipped in most articles about training load, so here it is plainly.

Damsted and colleagues ran ProjectRun21: 447 runners preparing for a half marathon, GPS-tracked, with injuries recorded prospectively. They tested ACWR against injury using multiple thresholds, coupled and uncoupled. They found no association. Not a weak one. None that held up.

Then there’s the paper that should have ended the conversation. Impellizzeri’s group calculated ACWR normally, then recalculated it replacing the actual chronic workload with random numbers, and found the random version was about as strongly associated with injury as the real one. Their title said the quiet part: time to dismiss ACWR. The mechanism is mathematical coupling. When you divide a value by an average that contains it, you manufacture correlation structure out of arithmetic rather than biology.

The 10% rule has the same problem. Buist and colleagues randomised 532 novice runners to a graded 13-week build versus a standard 8-week build and found roughly one in five runners got injured in both groups. The gentler progression didn’t help.

What survives all this is narrower but genuinely useful. Rasmus Nielsen’s work on novice runners found that those who increased weekly distance by more than 30% over a two-week window had elevated risk of the specific injuries you’d expect from distance progression: patellofemoral pain, tibial stress, iliotibial band syndrome, patellar tendinopathy. That’s a change-detection finding, not a ratio finding. Treat the acute chronic workload ratio the same way: as a change detector with a known bias, a way of asking “how different is this week from my recent normal,” not as a risk score with a red line.

Which means the useful workflow isn’t checking whether you’re at 1.4. It’s understanding why you’re at 1.4 and whether that particular reason is the kind your tissues care about.

Worked example: where a distance-based ratio lies to you

Here’s a real-shaped marathon build. Weekly distance, six weeks in:

Week 6 contains a 32 km long run with the final 8 km at marathon pace, and it reads 1.28. Inside the sweet spot. A dashboard would show green.

Now price the same weeks in session-RPE load (RPE 0 to 10 multiplied by session minutes, Foster’s method). Week 6 broken out:

Tue   10 km easy      55 min  RPE 3   =  165 AU
Wed   12 km, 5x2 km   66 min  RPE 7   =  462 AU
Thu    8 km easy      45 min  RPE 3   =  135 AU
Sat    8 km + strides 45 min  RPE 4   =  180 AU
Sun   32 km, last 8 MP 180 min RPE 8  = 1440 AU
                                        -------
                                        2382 AU

Weekly totals come out at 1150, 1320, 900 and 1485 AU for weeks 2 to 5. So chronic mean is (1320 + 900 + 1485 + 2382) / 4 = 1522, and the ratio is 2382 / 1522 = 1.57.

Same week. Same runner. 1.28 by distance, 1.57 by load. The distance figure missed the spike because that one Sunday was 46% of the week’s mileage and carried almost 60% of its load, and a weekly sum cannot see inside itself.

Uncoupling makes it starker. Drop week 6 from the denominator and distance gives 70 / 49.3 = 1.42 while load gives 2382 / 1235 = 1.93. Three numbers between 1.28 and 1.93 describing one week, all correctly calculated. That’s the reality of the acute chronic workload ratio in running, and it’s why the input choices matter more than the threshold.

Your choice of load unit decides your answer

Each metric is blind in a specific direction. Know which.

The Garmin blindness is worth dwelling on, because it’s the one most recreational runners are unknowingly relying on. Garmin Load is built from EPOC, a physiological strain estimate. A 3-hour easy long run and a 75-minute threshold session can land within spitting distance of each other on that scale. At 170 spm, the long run is about 30,600 ground contacts and the threshold session about 12,750. If what’s bothering you is a tibia or a plantar fascia, those two sessions are not remotely equivalent, and Garmin’s number cannot tell you so. Its “Strained” Training Status and its Acute Load Ratio band (a shaded optimal zone that in practice sits around 0.8 to 1.5 and shifts with your chronic load) are answering a cardiovascular question.

For most runners the pragmatic answer is to track two ratios: one on distance or step count, one on sRPE. When they disagree, the disagreement is the information.

Rolling averages hide timing; EWMA doesn’t

A 7-day rolling sum treats every day in the window identically. Your 32 km Sunday counts exactly as much on Monday morning as it does on the following Saturday, then vanishes completely overnight. That’s not how fatigue behaves.

Exponentially weighted moving averages fix the decay. The smoothing constant is λ = 2/(N+1), giving λ = 0.25 for a 7-day acute and λ = 0.069 for a 28-day chronic. Each day: EWMA_today = λ × load_today + (1 − λ) × EWMA_yesterday.

Run that 1440 AU long run against a runner sitting steady at 200 AU/day:

acute   = 0.25  × 1440 + 0.75  × 200 = 510
chronic = 0.069 × 1440 + 0.931 × 200 = 286
ratio   = 510 / 286 = 1.79   (was 1.00 the day before)

Take Monday off and it falls to 1.44, then keeps sliding. The signal arrives the day the load lands and fades the way soreness does. Strava’s Fitness & Freshness already works this way (42-day Fitness, 7-day Fatigue), as does TrainingPeaks CTL/ATL. Intervals.icu will compute EWMA ACWR directly and let you choose coupled or uncoupled and set your own windows. Runalyze does monotony and strain alongside it. If you want the step-by-step for pulling your own weekly figures out of your existing activity history, work through how to calculate your acute:chronic ratio from Strava and set the spreadsheet up once.

The layoff artifact, and why low chronic load breaks the ratio

Three weeks off with a chest infection. First week back you jog 20 km very easily. Coupled ACWR: 20 / ((0 + 0 + 0 + 20)/4) = 4.0.

Nothing is wrong. The ratio is garbage because the denominator is near zero, and this is exactly the artifact that drives much of the published association: low chronic load inflates the ratio, and people with low chronic load are often people who were recently injured or ill. The apparent risk signal is partly a returning-from-injury signal wearing a disguise.

The same distortion runs the other way at the top end. A ratio of 1.5 on top of 30 km/week means a 15 km jump. On 110 km/week it means 55 km. Identical number, entirely different demand. Always read the ratio next to the absolute figures, and ignore it outright for the first three weeks back from any break.

Split the ratio by tissue, not just by week

Different tissues respond to different things on different timescales, which is the strongest argument against a single whole-body number.

Bone responds to cumulative impact cycles, and its remodelling timeline is long. The resorption phase of bone turnover creates a window of weeks where the site is temporarily weaker than before the stimulus. A 28-day chronic window is too short to describe bone readiness: someone who did nothing for two months and has now put together four honest weeks looks fully adapted on a 28-day view and is not. Keep a 90-day rolling distance figure too. Intervals.icu will plot it.

Tendon cares about rate and magnitude of loading, so speed work, hills and plyometric volume drive Achilles and patellar tendon problems far more than gentle mileage does. Collagen turnover after a heavy tendon session is net-negative for around a day and a half, which is the physiological basis for the 48-hour minimum between hard sessions. Two quality days back to back is a tendon decision, not a fitness decision.

Muscle recovers fastest and forgives most, which is why you can feel fine and still be accumulating a problem elsewhere.

Practically: compute a separate ratio on your hard kilometres. Define the bucket as distance at threshold pace or faster, and track it as its own series. Take a typical Runna four-day marathon week: 12 km with 8×800 at 5K pace (6.4 km hard), 14 km with 3×3 km at threshold (9 km hard), 10 km easy, 26 km long with 10 km at marathon pace. That’s 62 km total with 15.4 km at threshold or faster, so 25% of weekly volume is hard. For most recreational runners, 10% is a sane ceiling and 12 to 15% is the experienced end. The total mileage in that week is unremarkable; the intensity fraction is where the plan is actually aggressive, and a whole-week ratio will never surface it.

The metrics that catch what ACWR misses

Monotony and strain. Foster’s monotony is the week’s mean daily load divided by the standard deviation of daily loads; strain is weekly load times monotony. Above 2.0 on monotony is the traditional flag. Run the numbers on the week-5 example above (daily loads 0, 165, 455, 135, 0, 180, 550): mean 212, SD 197, monotony 1.08. Now take a runner doing 10 km easy six days a week: mean 141, SD 58, monotony 2.45. The bland week scores worse on two-thirds the total load. That’s the whole case for hard days hard and easy days easy, expressed as arithmetic.

Long-run share. Keep the longest run under about 30% of weekly volume. Marathon builds break this rule by necessity in the last six weeks, which is precisely when you should be padding weekday volume rather than stretching the long run further.

Chronic load ramp rate. In TrainingPeaks terms, CTL rising 3 to 5 TSS/day per week is sustainable and 8+ is a gamble. This catches the plan that never spikes but never stops climbing either.

Consecutive-day structure. Count the gaps between hard sessions across the whole plan. One 36-hour gap in week 3 is a hole in the plan, not a detail.

HRV, resting HR, Garmin Training Readiness, Whoop recovery. These lag, they’re noisy at the individual-day level, and they’re confounded by alcohol, heat and sleep. Use them as confirmation rather than as triggers: three consecutive suppressed mornings alongside a ratio of 1.6 is a real conversation. One bad morning is a bad night’s sleep.

Auditing an AI plan before you run any of it

Export the plan into a spreadsheet and compute the ratios forward, for every week, before week one. This is the single highest-value thing in this piece and almost nobody does it. You have the whole plan in advance; you don’t need to wait and see.

Columns: week, total km, longest run km, longest run as % of week, threshold-or-faster km, hard km as % of week, projected sRPE load, coupled ACWR, uncoupled ACWR, gap in hours between each pair of hard sessions.

Then check five things.

Down weeks. Is there a 25 to 35% cutback every third or fourth week? ChatGPT-built plans routinely omit these, because “increase by 10% weekly” is a rule that compounds: 1.1^12 is 3.14, so a 12-week build from 40 km/week lands you at 126 km/week. Nobody writing the plan notices, and the plan won’t tell you.

Single-week jumps above 30%. Flag every one against Nielsen’s finding, and look hard at whether the jump is distance, intensity or both.

Both-at-once weeks. Any week that raises volume and raises the hard fraction is a week where you won’t be able to attribute what goes wrong. Change one variable per week.

The taper and the race. Race week ACWR drops well below 0.8 by design, and that’s correct. But check the two weeks after your goal race, because most plans just stop, and the resumption week often produces a 2.5 ratio against a collapsed denominator.

The long-run endgame. Read off the final six weeks specifically. A plan that runs 26, 28, 30, 32 km on consecutive weekends without a cutback is stacking bone load during the exact window when bone is mid-remodel.

Garmin Coach’s adaptive plans are conservative on volume and will pull sessions when Training Readiness tanks, so their failure mode is under-progression rather than spikes. Runna’s failure mode is the opposite: the intensity fraction. Plans you built with an LLM fail at structure, because the model produces arithmetically tidy progressions with no cutbacks and no sense of what fraction of the week a single session represents.

When the numbers say fine and your leg says otherwise

Symptoms override arithmetic every time. A ratio of 1.1 is not a permission slip.

Bone stress signals are the ones to take literally: pain you can locate under one fingertip, pain that gets worse as the run goes on rather than warming up, pain when hopping on that leg, ache at night. That sequence justifies stopping the plan and getting imaged, not a deload week.

Tendon behaves differently and is easier to manage. Morning stiffness that clears within 10 minutes is tolerable and often improves with loading. Stiffness lasting over half an hour, or pain that is worse 24 hours after a session than immediately after it, means the last session exceeded capacity. The workable rule is pain at or below 3/10 during the run, back to baseline within 24 hours, no progressive worsening week to week.

If a ratio comes back too high, resist the instinct to rest. Change one input. Hold total volume and convert one quality session to easy running, which drops sRPE load 15 to 20% without touching your chronic distance base. Or hold intensity and cap the long run at 25% of the week, redistributing the difference across two weekday runs. Insert a cutback at 70% of the previous week and recompute: in the six-week example above, making week 6 a 52 km week with a 24 km long run brings the load ratio from 1.57 to about 1.15, and the fitness cost of that swap is close to zero.

Log the four numbers weekly (distance, hard km, sRPE total, longest-run percentage) and you’ll have a chronic baseline worth computing ratios against by the time the plan gets serious.