Our Downhill Credit Was About Twice Too Generous. Here's the Fix.

On St. George's marathon course, net drop about 2,560 feet, our pace calculator used to credit an 8.6% faster finish from the downhill alone. On REVEL Mt. Charleston, one of the steepest net-downhill courses we cover, the credit was 17.3%. Both numbers came from the same treadmill-derived curve we've used for years, stretched well past where it was ever measured.

We searched 2.07 million marathon finishes across 95 races and matched 8,501 runners on the 8 downhill courses we could trace, comparing each runner's downhill-course time against their own flat marathons, weather- and altitude-adjusted. At St. George the real effect was 4.4% faster. At Mt. Charleston, 8.3%. Taken together, our old credit ran close to double the real thing, and a placebo run on 17 flat races came back at essentially zero, so the method wasn't manufacturing a downhill effect out of noise.

To fix it, we didn't fit a new curve straight to our own race results. Course by course, that data is too noisy to trust for a shape. Instead we started from a gentler downhill shape, the kind large-scale published running data has shown, then tested that shape against our own runners' race results on the 8 downhill courses we could trace. It held up well where the descent is steep.

Why the old downhill credit ran too hot

The grade cost underneath our splits and pace adjustments has always come from Minetti et al. 2002 (J Appl Physiol 93:1039): a fifth-order polynomial fit to the metabolic cost of running uphill and downhill, Cr = 155.4i⁵ − 30.4i⁴ − 43.3i³ + 46.3i² + 19.5i + 3.6 in J·kg⁻¹·m⁻¹, normalized to the flat-ground cost of 3.6.

It's a good study with a narrow purpose. Ten elite mountain runners ran 4-minute treadmill bouts with gas collected in the fourth minute, at 0% grade and then only at ±10% to ±45%. Nothing was measured between 0% and 10%, which is most of a road marathon: the −5% to +5% range where a course spends most of its miles is the polynomial's interpolation, not a measurement. A 4-minute bout with steady-state gas exchange also can't see what a 20-minute descent does to your quads later in the race. Minetti's own paper notes that the model overestimates downhill race speeds.

The metabolic cost itself still holds up. A 2022 treadmill study (Whiting, Hoogkamer & Kram) measured about 0.76 of flat-ground cost at −5.2% grade, close to what the polynomial predicts. What over-credited runners wasn't the energy cost. It was using that same curve to price how much faster a downhill course lets you race, not how much oxygen it takes to run one.

Measuring it against race results

We compared 2.07 million marathon finishes from 2008 on, across 95 races including St. George. The design is within-runner: each downhill-marathon finish is compared against the same runner's median time across their own flat marathons within about a year either side, never against the field.

• Matching. Strict name, sex, and age consistent with the date gap (within 1.5 years), plus state where the results carried one. A name had to be unique within a race-year, and any conflicting namesake dropped that result rather than risk a false match. Match rates ran 23 to 57% of eligible finishers depending on the race.
• Good days and bad days. Each runner's flat baseline is the median of their own flat marathons. When a runner had three or more flats, any more than 25% off their own median was dropped as an outlier day, and ratios more than 25% from the race median were trimmed. The result held up at 15% and 35% thresholds instead, and with Huber and trimmed means in place of medians. Confidence intervals are bootstrapped over runners, 1,000 draws.
• Training drift. Finish times drift for reasons that have nothing to do with grade, so we checked whether the pairs were mostly people running their flat marathon before or after their downhill one. Both orders showed up in roughly balanced numbers. Drift measured out at −0.38% per year (±0.11), and it cancels in a within-runner design. A 6-month window instead of a year changed nothing.
• Weather and altitude. Every time went through our own race-day weather model: WBGT with the pace-conditioned heat slope, temperature-ramped rain, and for the big-drop courses, weather blended from start to finish by race progress rather than one number for the whole field. Wind was left unadjusted. Altitude was adjusted with our altitude model on each course's effective elevation.
• Placebo. We ran the same pipeline on 17 flat races, telling it to treat them as if they were downhill. The average effect it found was +0.01%, with about ±1 point of per-race noise (SD 1.08%). That's the floor: anything the real downhill races show below about a point isn't distinguishable from this noise.

What we found

Eight downhill courses had course traces we could grade mile by mile and enough matched runners to test a curve against. St. George came in at 4.4% faster [95% CI −4.7, −4.1], against our old credit of 8.6%. REVEL Mt. Charleston, the steepest course in the set, came in at 8.3% faster against our old 17.3%. Two of the eight, Mesa and Napa, came in essentially flat to slightly slower than their own runners' flat marathons: courses the old model still credited 3.1% and 1.0%.

Fit across all eight: the old model's predictions explain the observed effect at a slope of 0.48, meaning for every point of real downhill benefit it credited about two. Mean absolute error 3.20 points, RMSE 4.00 (4.03 weighting each race by its matched-runner count, so the bigger backtests count more). The new curve fits at a slope of 1.07, MAE 1.24, RMSE 1.40 (1.71 weighted). A uniformly halved Minetti curve (multiply the whole credit by 0.45) fits about as well on these particular courses. The data can't separate the two shapes below about −6% grade. Where they do differ is on steeper grades, which is where the new curve's flattening matches the shape large-scale published running data has shown.

The new curve

Downhill pace cost is now cost = 1 + 2.70·g + 15.3·g² for grade g below zero (g as a decimal, so −0.05 for a −5% grade), evaluated per mile. Uphill is untouched: it's still the Minetti polynomial. The downhill benefit bottoms out near −9% grade at about 0.88 of flat pace cost, and it's gone by around −18%, where the credit is held at flat rather than letting it reverse into a penalty.

The shape we started from is a gentler downhill curve of the kind large-scale published running data has shown. Our version reads 0.95 of flat pace cost at −2% grade, 0.90 at −5%, and 0.88 at −10%. Our race-results test is independent of where that shape came from: it checks a curve like this against real finish times on 8 downhill courses, not more training-run data, and it holds up where the descent is steep.

Other signals point the same way. Running with Rock's analysis of Boston qualifiers found steep-downhill qualifying courses running about 2 to 2.5% faster than flat ones, and REVEL-style courses about 3.5 to 5%, both closer to our new curve than the old one. Starting in 2027 the B.A.A. will add 5:00 to qualifying times run on courses that drop 1,500 to 2,999 feet, and 10:00 for 3,000 to 5,999 feet, roughly a 2.5 to 3% adjustment. World Athletics caps the net drop allowed on a record-eligible course at 1 meter per kilometer (Rule 31.21.3), which exists because event organizers already know a downhill course is fast enough to need a rule about it. For background on the shape of an equal-effort downhill pace curve, see Strava Engineering's "An Improved GAP Model" (Robb, 2017), built from about 6 million runs.

Who this does and doesn't depend on

We checked whether the downhill benefit depends on how fast you are. It mostly doesn't. Split by quartile relative to each race's own median finish: the fastest quarter gained 0.01 points more than predicted, the second quarter 0.37 less, the third 0.11 more, and the slowest quarter (4:18 to 6:00 marathons) 0.45 more [95% CI −0.04, +0.91]. That interval crosses zero, so speed makes no reliable difference here, and nothing strong enough to build a pace-dependent adjustment on.

Boston, +3.4% versus the same runners' flat marathons, isn't in the fit above. It's not a grade test: Boston selects for qualifiers, has the Newton hills working against a net-downhill profile, and draws a lot of tourist pacing. Folding it in would mean fitting a downhill curve to a course that isn't testing the downhill effect in isolation.

Grade also has to be measured at the right resolution. We compute the effect mile by mile, not on raw trace segments. On Super Hyak, per-segment grades on a noisy course trace produced a fake average cost of +0.2% per segment, while the same course measured per mile showed the −4.2% its profile actually has. Small trace segments can manufacture cost on a flat course and erase real downhill credit on a hilly one. Whatever curve you use, the grades feeding it have to come from a sustained window, not a raw GPS point cloud.

Why you might still run slower than the model says

None of this prices what a long, unfamiliar descent does to your legs. Steady-state gas exchange, the kind Minetti's study measured, can't see eccentric quad damage: the muscle-fiber strain from absorbing your body weight over and over on the way down. That damage raises the metabolic cost of running later in the race. One study found oxygen cost on flat ground up about 18% right after 30 minutes at −15% grade (Lima et al. 2021), and the elevated cost can persist for days afterward (Chen, Nosaka & Tu 2007; Braun & Dutto 2003), not just for the rest of that run (Bontemps et al. 2020 review, Sports Med).

Training protects against this. The repeated-bout effect, described by Byrnes et al. (1985) and confirmed by Eston et al. (2000), means quads that have handled downhill running before absorb a big descent with much less damage the next time. That's also why the pace calculator carries a caveat instead of a bigger deduction: for a descent where hills help by 2% or more, or the net drop is 1,500 feet or more, the finish-time card notes that the number assumes trained quads. Signed in with synced training, we check your own recent descent running against the course's demand and swap in a line that says so, on courses with a trace.

Also changed: altitude

The same pass added altitude to the race-page pace calculator: about 1% per 1,000 feet above 2,000 feet, on the course's effective altitude (the mean elevation penalty along the trace, not just the start line). That's deliberately conservative. St. George's effective altitude on the current course trace is about 4,190 feet, where our model prices it at 2.2% slower. The aerobic-power literature closest to Minetti's own paper puts an altitude near 4,000 feet at about 2.2% slower; Péronnet, Thibault & Cousineau's 1991 marathon model puts it close to 3.7%. It's a small addition next to the downhill change, so that's the whole story here for now.

Where you'll see it

• The race-page pace calculator: Race-day finish, True Effort, and the Hills line in the effect ledger
• Mile-by-mile terrain splits
• The course explorer's effort map
• The standalone pace calculator
• Training-run pricing on synced runs

What changes on the site

Course hills figures under the new model, as the pace calculator computes them mile by mile on each course's own trace:

• St. George: 3.5% faster (was 8.6%)
• Boston: no effect (was 1.2% faster), and still not a grade test in its own right
• London: 0.2% slower (was 0.9% slower). The old figure was mostly elevation noise on a flat course, priced segment by segment.
• Berlin and Erie: essentially unchanged (0.2% and 0.1% slower)

These are the live calculator's numbers. The backtest charts earlier in this post show the new curve's fitted prediction for each of the eight courses instead, which can read a little different. St. George's backtest prediction, for example, is 3.7% against the 3.5% the calculator shows today.

More guides

• Why We Rebuilt RunScore From 6.3 Million Marathon Finishes — the same within-runner approach, applied to heat, rain, and wind
• How Weather Affects Your Marathon Pace — the rest of the pace-adjustment model this curve slots into
• Marathon Pacing Guide — adjusting effort and splits when the terrain or the weather turns against you