Sectional Efficiency Part 2 - Putting Numbers On it

HK Speed King
Aug 14
9 min read

Updated: Aug 15

HK SPEED KING — SECTIONAL EFFICIENCY PART 2

First Published: 14th August, 2026

Part 1 made a case in words: the final 600m of a race is the mathematical remainder of the energy spent earlier, so a fast closing split in a slowly-run race is worth very little. That argument is easy to nod along to and hard to act on. Nodding along doesn't tell you ⟨how much⟩ to discount a fast closing split, or whether the correction is large enough to change which horse you back.

So let's put numbers on it.

Everything below comes from a small, deliberately transparent sample: 84 runs by three horses at Sha Tin and Happy Valley between November 2022 and July 2026, all on turf. The three were chosen because they sit at opposite ends of the ability scale and therefore make the arithmetic visible:

Horse	Runs (wins)	Median SP	Grade Range
Ka Ying Rising	22 (20)	$1.10	Class 4 to Group One
A Americ Te Specso	54 (7)	$14.0	Class 4 / Class 3
Le Zonda	8 (2)	$18.50	Class 2 / Class 3

Ka Ying Rising's 22 starts were all at Sha Tin, A Americ Te Specso raced 49 of 54 times at Happy Valley, and Le Zonda split his eight between the two. That venue split matters later, and we come back to it.

This is a worked illustration, not a validation study. Eighty-four runs settles nothing on its own. What it does is make a mechanism visible at a scale you can check by hand — and every effect shown here is one we can then go and test properly on the full database.

Defining the metric precisely

Part 1 spoke loosely about "the final 600m". Hong Kong sectionals don't come in 600m units, so let's be exact. HKJC publishes splits in 400m blocks measured back from the winning post, with any odd remainder falling in the opening section:

Distance	Section Structure
1000	200.400.400
1200	400.400.400
1400	200.400.400.400
1600	400.400.400.400
1650	450.400.400.400
1800	200.400.400.400.400
2000	400.400.400.400.400
2200	200.400.400.400.400.400
2400	400.400.400.400.400.400

The final section is always a clean 400m, so that is what we'll use.

The standard industry measure is finishing speed percentage:

Finishing speed % = (average speed over the final 400m) ÷ (average speed over the whole race) × 100

Above 100% means the horse finished faster than its own race average.

Below 100% means it slowed. It is the most common sectional metric in circulation, and it is the one we are about to break.

Note the definitional problem hiding in it: the denominator contains the early part of the race. Run faster early and you raise your own race average, which ⟨mechanically⟩ lowers your finishing speed percentage even if the final 400m is unchanged. Any anti-correlation between early pace and finishing speed % is therefore part physiology and part arithmetic. We'll separate those two later, because most published sectional commentary does not.

The test that should worry you 😟

Here is the median raw finishing speed percentage for each of our three horses, across every run in the sample:

Horse	Median SP	Median Raw Finishing Speed %
Le Zonda	$18.50	102.51%
A Americ Te Specso	$14.00	102.42%
Ka Ying Rising	$1.10	102.36%

Ka Ying Rising, a horse who won 20 of 22 starts including multiple Group 1s, and who the market priced at odds-on almost every time he ran, produces the lowest figure of the three — and all three sit inside a range of 0.15 percentage points.

Raw finishing speed percentage cannot tell a Group 1 sprinter apart from a Class 3 handicapper. Not "does it imperfectly". Cannot.

That is not a flaw in the horses. It is the metric measuring the wrong thing. Ka Ying Rising's races were run fast in front of him and by him; the handicappers' races were often crawls. The metric reports the race, not the horse.

How much of the variance is pace, not merit?

Standardise every run's early pace against the median for that track and distance, then correlate it with finishing speed percentage:

r = −0.47 across all 84 runs

r = −0.62 once finishing speed % is also standardised within track and distance

Squared, that second figure means roughly 38% of all variation in raw finishing speed percentage is explained by early tempo alone — before we know anything whatsoever about the horse.

Now the honest part, and the step most sectional commentary skips.

How much of that is real physiology and how much is the shared-denominator artefact flagged above?

The test is a shuffle control. Keep every run's actual final 400m time, but randomly reassign which early split it is paired with, within the same track and distance. That destroys any genuine relationship while preserving the arithmetic. Repeat 2,000 times:

Measure with early pace	Correlation
Observed	-0.62
Shuffle Control (arithmetic only)	-0.44 (95% range -0.59 to -0.26)
Genuine excess	-0.18

So roughly two-thirds of the effect is definitional and one-third is real. Both matter for handicapping — the definitional part is exactly why the metric fails as a merit measure — but anyone claiming a fast closing split "proves" a horse was hard done by is leaning on a number that is largely an accounting identity.

For the physiological half, use a measure with no shared denominator at all: early pace against raw final 400m time in seconds.

r = +0.31. The faster the early pace, the slower the final 400m. Each additional standard deviation of early pace costs roughly +0.11 seconds on the closing split.

Positive correlation, no arithmetic contamination. The energy toll Part 1 described in prose is measurable, and in this sample it is worth about a tenth of a second per standard deviation of early pressure.

Fatigue resistance is a horse-level parameter

Run that same clean test ⟨within⟩ each horse. This is the strongest evidence available here, because comparing a horse to itself controls for ability, class and constitution entirely:

Horse	Runs	Seconds added to final 400m per 1 SD faster
Le Zonda	8	0.28s
A Americ Te Specso	54	0.19s
Ka Ying Rising	22	0.08s

A Americ Te Specso, over 54 runs, reliably runs a slower closing split when it goes harder early. The pattern is internal to one horse and cannot be explained by class, venue or opposition.

And the elite horse pays roughly a third of the penalty the handicappers pay for the same relative increase in early pressure. That is what "class" looks like when you measure it in seconds: not a bigger sprint, but a smaller bill for the sprint.

Treat that specific comparison as a hypothesis rather than a finding — 22 runs against 54, and a horse who was rarely under genuine pressure because he was usually in front. But it is precisely the kind of horse-level parameter a sectional model should be estimating, rather than assuming everyone decelerates at the same rate.

Two traps that survive even careful analysis

The section-length trap. A common shortcut is comparing opening-section times across races. Don't. Look again at the section structuretable, the opening section is 200m at 1000m and 1800m, 400m at 1200m and 2000m, and 450m at Happy Valley's 1650m. An identical honest tempo produces wildly different opening ⟨times⟩ purely because the sections are different lengths. Any pace ratio built from raw section times partly encodes which distance family the race belongs to rather than how it was run. Convert to metres per second before comparing anything across distances — always.

The run-up contamination. The run to the first turn varies enormously by course, distance and rail position. A long run-up means the field is still winding up through section one; a short one means immediate compression. That is a fact about the track, known before the race is run, and it has no business being read as a fact about the horse.

The correction, and whether it works

The fix: if early tempo explains a large share of finishing speed percentage, remove that share and keep the remainder. Regress finishing speed % on standardised early pace within track and distance, then take each run's residual — how much faster the horse finished than a run at that tempo would predict.

This is a deliberately simplified stand-in for what the H K Speed King Sectional Efficiency Index does in the engine. The production version carries class pars, surface weighting, field-level pace context and horse-specific terms. The principle is the same: score the deviation, not the raw number.

Apply it to the three horses:

Pace-adjusted figures provide more insight

The metric that could not separate them now ranks them correctly, in the order the market and the results independently agree on.

But ranking three horses you already know the answer for proves nothing. The real test is whether the adjustment predicts ⟨within⟩ a horse — which of A Americ Te Specso's own 54 runs were the good ones. Scoring each run against whether it produced a top-three finish:

0.50 is a coin flip. Raw finishing speed percentage is barely better than one. The adjustment roughly doubles the signal over raw finishing speed % — on one horse, 54 runs, so wide error bars, but the direction is unambiguous and it is measured on the population that matters.

The worked example that makes it concrete

Two Ka Ying Rising runs at Sha Tin, 1200m:

Read the sectional line alone and the 2024 run is the good one — 104% finishing speed, a genuinely eye-catching figure. The 2025 run finished ⟨below⟩ its own race average and looks like a horse going backwards.

The 2025 run was 1.49 seconds faster overall, in a higher grade, and won. It carries the fastest early split of the horse's 22-run sample. He didn't slow down because he was beaten; he slowed relative to his average because he had already gone faster than anyone else could early. The "worse" sectional profile is the better performance by a distance.

The trajectory over his twenty 1200m starts says the same thing:

He got faster early ⟨and⟩ faster late simultaneously. Raw finishing speed % registers part of that improvement. The pace-adjusted figure registers the scale of it.

And the flawed flier, quantified

Take the six highest raw finishing speed percentages in the entire sample. If the metric measured merit, these should be the best runs in the dataset.

They produced two wins from six, at a median starting price of $17. The single highest figure of all 84 runs — Le Zonda's 107.63% at Happy Valley over 1800m in March 2026 — finished fourth, at $52.

All six came from Happy Valley staying races, and five of the six were run at least a full standard deviation below the median early tempo for their track and distance.

The metric wasn't finding good horses. It was finding slow first halves.

Where this leaves you

Nothing here overturns Part 1 — it puts sizes on it, and the sizes are larger than most people assume:

Roughly 38% of raw finishing speed % variation is early tempo, before any horse-specific information enters.
About two-thirds of that is definitional, an artefact of dividing by an average that contains the early sections.
The genuine physiological cost is about +0.11s on the final 400m per standard deviation of early pace — and that rate differs by horse, which is the part worth modelling.
Correcting for it roughly doubled the within-horse discrimination of the metric in this sample.

Revised checklist

✅ Convert to speeds before comparing anything across distances. Section lengths differ; raw times are not comparable.

✅ Never compare finishing speed % between races without pace-standardising first. Between-race comparison of the raw figure is close to meaningless.

✅ Ask what the metric shares a denominator with. If early pace is in the denominator, some of your "signal" is arithmetic.

✅ Discount the opening section by course and rail, not by intuition — the run to the first turn is a track fact, not a horse fact.

✅ Score the residual, not the number. How much better than the tempo predicted, not how fast the clock said.

✅ Expect the best horses to look ordinary on raw sectionals. They are usually the reason the sectionals look that way.

The three horses above have wildly different ability and nearly identical raw sectional signatures. If your form study stops at the closing split, those three are indistinguishable to you — and the market has already worked out which is which.

A note on sample size

Eighty-four runs across three horses is an illustration, not a validation. Two of the effects shown rest on a single horse, and Ka Ying Rising raced exclusively at Sha Tin while A Americ Te Specso raced mostly at Happy Valley, so the between-horse comparison carries a venue confound that the within-horse tests do not. We publish the mechanism and the method here; the parameter estimates that go into the report come from the full database, not from three horses. Treat every figure above as a demonstration of ⟨how the correction behaves⟩, not as a calibrated coefficient.

References:

Benter, W. (1994). Computer based horse race handicapping and wagering systems: A report. In D. B. Hausch, V. S. Y. Lo, & W. T. Ziemba (Eds.), Efficiency of racetrack betting markets

https://gwern.net/doc/statistics/decision/1994-benter.pdf