Model validation

Checking Lukra against the textbook

We re-derived the milk a cow should make from her feed using two independent industry standards — NRC 2001 and NASEM 2021 — and compared them to what Lukra predicts. Across nine New Zealand scenarios spanning three breeds and both parities, Lukra lands within a few percent of one of the two standards in every run.

Scope 9 runs · 3 breeds · parity 1 & 2 Benchmarks NRC 2001 · NASEM 2021 Basis energy (NEL) partitioning Workbook NRC2001_LukraValidationSpreadsheet_12.xlsx

Nine runs, two standards

Lukra vs the nearer benchmark
250300350400450kg milk solids / cow / lactationFoxton, Lower Nth Is.−4.2% vs 2021DairyMax 2nd farm−2.9% vs 2001Waikato, Horsham Downs+2.2% vs 2021Jersey, single cow+0.5% vs 2001Jersey herd (4000)+3.0% vs 2021Jersey heifers (4000)−5.0% vs 2001Kiwi Cross, single cow−2.0% vs 2001Kiwi Cross herd (4000)−2.2% vs 2001Kiwi Cross heifers (4000)−3.0% vs 2001
NRC 2001 NASEM 2021 Lukra

Each grey bar is the gap between the two textbook standards for that herd; the green diamond is Lukra, and the figure on the right is its distance to the nearer standard. Across all nine runs that distance is a few percent — Lukra lands within or close to the band between the two editions, mature cows and heifers alike.

Why this matters

Lukra derives the ration that makes a pasture-based dairy herd the most money, month by month across a season. It uses the equations derived by Hulme et al. (1986) and energy partitioning rules so that the dairy-cow model and its associated Linear Program can optimise a year-round feed budget that takes both economics and biology into account.

But a profit number is only as trustworthy as the biology underneath it — so before anyone leans on Lukra's margin, the milk it predicts from a given feed supply has to stand up against the published science. So we asked ourselves the question: when Lukra formulates a feed budget and reports energy and protein intake, calculated body condition score and milk-solids production, what milk-solids production does one of the gold standards (NRC 2001 and 2021) predict for the same animal type — and how does that compare with the MS production predicted by Lukra?

This is an account of where Lukra agrees with the standard energy accounting, where it differs, and exactly why it differs.

The short version Lukra's milk prediction agrees with modern energy standards to within a few percent across all nine scenarios we ran — mature cows and heifers alike. That agreement is itself a check on the mechanism: it says the energy partitioning between milk and body fat inside Lukra's linear program is behaving the way the textbook accounting says it should.

What we checked it against

Does a cow's milk come from a response curve, or from where her energy is sent once the essentials are paid for?

The industry reference for "how much milk can this feed support" is the National Research Council's energy-allowable milk. The idea is simple bookkeeping: take the energy in the feed, subtract what the cow needs for maintenance, walking and grazing, pregnancy and growth, and convert whatever is left into milk. We computed this two ways — the 2001 edition (NRC) and the 2021 revision (NASEM) — entirely independently of Lukra, from the same monthly feed intakes and animal types that were used in the Lukra optimisation.

One subtlety matters for everything that follows. Energy-allowable milk in NRC has no yield ceiling: it assumes every spare megajoule above the essentials becomes milk, and that body reserves never change. A real cow doesn't behave that way. Past her peak, or while she's rebuilding condition, she banks surplus energy as body fat rather than pushing it all into the udder. That is why Lukra models increasing fat deposition as the cow approaches her genetically maximum production at any stage of lactation. So the standard (NRC) is best read as an uncapped energy-to-milk conversion — useful precisely because any gap against it is informative.

Lukra carries the biology the conversion leaves out

Lukra starts from a lactation curve that sets each cow's milk potential for the month. Energy is turned into milk up to that potential — and any surplus beyond it is partitioned to body fat, not forced into milk. That partitioning is the job the linear program does: after maintenance, growth and pregnancy are met, the solver decides how the remaining energy splits between milk and reserves, choosing whatever maximises season-long margin. It's the same most-limiting-nutrient logic NRC uses for the essentials, with one honest addition — a real cow can say "no more milk this month" and put the rest on her back.

Context — the redesign this validates Before beta testing began, Lukra went through a redesign of how it prices the most productive part of a cow's output. This validation is the check that the redesign holds up — across breeds, herds and both parities, and against a newer standard than the one it was originally checked against.
Go deeper: what the earlier version got wrong for beta-testers

An earlier, pre-beta version of Lukra read the Hulme cost as a rising price on milk itself. In effect, the "expensive" portion of high production was priced away rather than redirected anywhere — body fat effectively went missing from the ledger, and the implied efficiency of that marginal milk was extremely low. The result was a mid-lactation under-prediction of milk solids by as much as 30 % against NRC 2001.

The redesign keeps the underlying Hulme equations but re-attributes that expensive portion explicitly to body fat — the partitioning described above — which closed the gap to under 2 % for a mature cow. This post is the follow-up question: does that fix hold up across many herds, breeds and parities, and against the newer 2021 standard?

Mature cows: near-coincident

Start with the clean case — a cow at or near her mature weight, where milk really is the energy-limited output and both models should agree. Here is a Kiwi Cross cow across the full lactation, the same monthly feed entered into all three calculations:

0.60.81.01.21.41.61.8kg MS / cow / dayJulAugSepOctNovDecJanFebMarApr
Lukra NRC 2001 NASEM 2021

Lukra tracks NRC 2001 almost exactly through the productive lactation and sits just above NASEM 2021. Full-lactation totals: Lukra 395, NRC 2001 403 (−2.0 %), NASEM 2021 365 (+8.2 %).

This is the result that matters most, and it's quietly remarkable: Lukra uses a maintenance cost (Agnew–Yan) that is neither NRC value, yet it lands between the two editions and hugs the line through the months that actually make the milk. The energy accounting is sound.

Go deeper: why NASEM 2021 is the tougher benchmark for beta-testers

The 2021 revision changed two coefficients that pull in the same direction. It raised maintenance from 0.080 to 0.100 × BW^0.75 (+25 %), and lifted the ME→NEL efficiency from 0.64 to 0.66. The net effect is that 2021's allowable-milk line sits below 2001's — it's the more demanding standard. Lukra sits with the modern line on maintenance-driven terms while landing nearer 2001 on the totals, which is exactly what you'd expect from a model whose maintenance is independently derived rather than copied from either edition.

All nine runs

Repeating that across the herd set: three breeds, single-cow and 4000-cow scenarios, two parities. The bolded residual in each row is the gap to the nearer of the two standards — the honest "how far from the textbook" number.

Full-lactation milk solids (kg / cow). Residual = (Lukra − standard) ÷ standard.
ScenarioBreedPar.LukraNRC 2001NASEM 2021vs 2001vs 2021
Foxton, Lower Nth Is.Holstein-Friesian2288343301−15.9%−4.2%
DairyMax 2nd farmHolstein-Friesian2345356313−2.9%+10.2%
Waikato, Horsham DownsHolstein-Friesian2339374331−9.4%+2.2%
Jersey, single cowJersey2363362326+0.5%+11.5%
Jersey herd (4000)Jersey2421440409−4.4%+3.0%
Jersey heifers (4000)Jersey2349367312−5.0%+12.0%
Kiwi Cross, single cowKiwi Cross2395403365−2.0%+8.2%
Kiwi Cross herd (4000)Kiwi Cross2363371332−2.2%+9.3%
Kiwi Cross heifers (4000)Kiwi Cross1294303239−3.0%+23.2%

All nine sit within roughly 5 % of their nearer benchmark. Mature cows cluster tight to NRC 2001; the herd-scale runs behave just like their single-cow counterparts, so nothing odd creeps in at scale — heifers included.

Go deeper: a calculation check the process caught for beta-testers

The two heifer scenarios (Kiwi Cross and Jersey, both 4000-head) initially showed the largest gaps in the set — Lukra running a few percent above NRC 2001. Tracing it down, the gap wasn't in Lukra: the comparison spreadsheet was charging maintenance on the animal's mature bodyweight rather than her actual, lighter, growing bodyweight. For a mature cow the two are nearly identical, so the error was invisible everywhere else in the set — it only shows up for a growing animal.

Correcting the maintenance calculation to use actual bodyweight brought both heifer scenarios in line with the rest of the set, within a few percent of NRC 2001 like every other run. If anything, this strengthens the result: it's exactly the kind of check this validation exercise exists to catch, and the milk/body-fat partitioning inside Lukra needed no adjustment at all.

Every run, month by month For beta-testers

The totals in the table are the area under these curves. Here is each of the nine runs as its own panel — Lukra against both standards, on a shared scale so the levels are comparable. It's the detail behind the overview chart at the top: all nine panels show the three lines running close together through the season, heifers included.

Lukra NRC 2001 NASEM 2021
Foxton, Lower Nth Is.−4.2% vs 2021
12JulDecApr
DairyMax 2nd farm−2.9% vs 2001
12JulDecApr
Waikato, Horsham Downs+2.2% vs 2021
12JulDecApr
Jersey, single cow+0.5% vs 2001
12JulDecApr
Jersey herd (4000)+3.0% vs 2021
12JulDecApr
Jersey heifers (4000)−5.0% vs 2001
12JulDecApr
Kiwi Cross, single cow−2.0% vs 2001
12JulDecApr
Kiwi Cross herd (4000)−2.2% vs 2001
12JulDecApr
Kiwi Cross heifers (4000)−3.0% vs 2001
12JulDecApr

Milk solids, kg/cow/day, Jul→Apr. Shared y-axis 0–2.1. The first-month dips on some herd-scale panels (e.g. Jersey and Kiwi Cross herds) reflect the LP choosing not to feed to full animal potential that month — it isn't obliged to, and won't, if the energy pays off better spent elsewhere in the season. Residual shown is to the nearer standard.

What this check does and doesn't cover

  • Energy only. The comparison assumes protein is adequate and energy is the limiting nutrient — reasonable for NZ pasture. Lukra's full protein and amino-acid modelling isn't part of this validation.
  • The 2021 grazing term is held equal to 2001 in our 2021 column (the exact 2021 coefficient wasn't pinned down). On pasture the effect is small, and the cell is editable in the workbook.
  • One scenario family per breed. Nine runs is a solid spread across breeds and parities, not a population. The pattern is consistent enough to trust the mechanism; a wider batch would tighten the late-lactation behaviour further.
Go deeper: the growth-efficiency caveat optional

Lukra uses a 0.62 ME→NE efficiency for growth in the lactation context (not the ~0.40 feedlot figure). We never had to take that on faith: the independent growth calculation — which doesn't use 0.62 at all — matches Lukra's retained energy of gain to within ~3 %, so the growth side is validated from the outside.

The verdict

The energy accounting holds — across every scenario we ran, not only the mature cows. Lukra tracks the modern standards within a few percent across three breeds, both parities, and single-cow through 4000-head herds. The two heifer scenarios initially showed a larger gap; tracing it down led to a maintenance-calculation issue in the comparison spreadsheet, not in Lukra, and correcting it brought both runs in line with the rest of the set. Nothing in the nine runs points to a partitioning error.

  • All nine runs track the standards to within a few percent of the nearer benchmark, across three breeds and both parities.
  • Herd scale matches single-cow runs — no surprises at 4000 cows.
  • The milk/body-fat partitioning holds for growing heifers too, once a maintenance-calculation issue in the comparison spreadsheet was corrected.

New to Lukra? Start with the overview: what it is and why it plans for profit

References & sources

NRC (2001). Nutrient Requirements of Dairy Cattle, 7th rev. ed. National Research Council. Energy Ch. 3, Protein Ch. 6.
NASEM (2021). Nutrient Requirements of Dairy Cattle, 2021 revision. Raised maintenance (0.080→0.100 ×BW^0.75) and ME→NEL efficiency (0.64→0.66).
Agnew & Yan. Maintenance energy basis used by Lukra (0.57 × metabolic BW, ME) — independent of either NRC value.
Hulme, Kellaway & Booth (1986). The CAMDAIRY model. Agricultural Systems 22: 81–108. (The milk-response equations Lukra builds on.)
Daniel, Friggens et al. (2016, 2017). Milk response to energy and protein; cow potential by stage of lactation. Animal 10; J. Dairy Sci. 100.
Friggens, Ingvartsen & Emmans (2004). Prediction of body lipid change. J. Dairy Sci. 87: 988–1000.
Macdonald et al. (2008). NZ stocking-rate trial. J. Dairy Sci. 91: 2151. (Breed averages, MS/cow benchmarks.)

Reproduced from NRC2001_LukraValidationSpreadsheet_12.xlsx, ScenarioLog tab — nine frozen runs. Comparison is energy-only; agreement within a few percent across all nine runs, mature cows and heifers alike. Lukra — Smarter feeding. Greater profit.