One thing that I try to do is go back and revisit previous analysis to make sure it still holds up. I’ve spent a lot of time over the past few years using simulated combat to ask a fairly simple Kings of War question: which units give us the most fighting power for their points?
The latest analysis turned up a result that initially looked dramatic. In an earlier pass, all 77 matched regiment-to-horde upgrades lost ground when I divided their Elo rating by their points cost. That sounds like evidence of a horde tax. The problem is that the calculation itself has a flaw. Once I changed how I measured efficiency, the story became much more interesting–and less definitive.
TL;DR: The apparent horde tax depends on the yardstick. A better value measure makes regiment-to-horde upgrades look close to neutral relative to a size-adjusted baseline.
First, what does the simulator actually measure?
The current dataset contains 725 simulated unit profiles from 20 factions, including 574 troops, regiments, and hordes. Another 232 Individual-sized profiles were excluded from the Elo calculations. The analysis uses the 4.1 unit catalogue and a simulation run of a bit more than 7 million combats between units. The model runs units through one-on-one fights, then assigns Elo ratings. If you’ve seen Elo in chess, the idea is familiar: beating a stronger opponent earns a bigger rating gain than beating a weak one. Higher Elo means a unit performed better in these simulated duels.
There is an important limit here. A horde with 2,000 Elo is not a 2,000-point unit, and a 200-point unit should generally beat a 100-point unit. I want to know whether a unit is strong for what I paid, not just whether it wins a straight fight.
Earlier versions of my unit-efficiency analysis used Elo per point as one measure. Unfortunately, as I’ve thought through this more, I don’t think the analysis holds up: Elo doesn’t have a meaningful zero. Elo doesn’t scale linearly like points do, so the ratio between units shifts as the baseline shifts. Imagine a 100-point unit rated 1,000 Elo. That’s 10 Elo per point. Add 500 to every unit’s Elo rating without changing a single combat result, and it becomes 15 Elo per point. The battlefield hasn’t changed, but the supposed efficiency has. That’s a problem if we’re going to make claims about unit pricing.
A better measure, with an important catch
The revised model begins by estimating what Elo a unit would normally have given its points cost and size. In this dataset, an additional point of cost corresponds to about 5.55 Elo, after adjusting for size. From there I calculate three related measures:
- Elo above expectation: the gap between a unit’s actual Elo and the rating predicted from its cost and size. Positive means it exceeded that fitted baseline.
- Implied points: its listed points, plus the Elo gap converted back into points using that 5.55 Elo-per-point slope.
- Points Value Ratio (PVR): implied points divided by listed points. A PVR of 1.20 means the profile’s rating maps to a price 20% higher on the model’s same-size relationship.
For readers who like the math:
Implied points = listed points + (actual Elo − expected Elo) / 5.546
PVR = implied points / listed points
The revised combined cost-effectiveness ranking puts equal weight on standardized PVR and standardized Elo above expectation. Those are two views of the same underlying Elo gap, so I don’t treat the combined ranking as a separate piece of evidence. The dataset also records an Elo-implied expected score versus par: a theoretical matchup with a typical unit of the same cost and size, not an observed tabletop win rate. PVR fixes the arbitrary-zero problem. If every Elo rating rises by 500, PVR stays the same. The catch is the formation-size adjustment. That matters a great deal for the horde question.
So what happened to the horde tax?
I matched the same unit family within a faction across sizes, keeping weapon variants distinct. The table reports how many model-implied points the upgrade gained per extra point paid.
| Upgrade | Matched families | Median extra cost | Median adjusted return | Returns above 1.00 |
|---|---|---|---|---|
| Troop → Regiment | 215 | 40 pts | 1.01 | 114 / 215 (53%) |
| Regiment → Horde | 76 | 75 pts | 0.99 | 37 / 76 (49%) |
| Troop → Horde | 63 | 105 pts | 1.05 | 41 / 63 (65%) |
At first glance, the result looks almost perfectly balanced. The typical regiment-to-horde upgrade returns about 99 cents on the model-implied dollar. Compare that with a median 11% decline in Elo per point for the same upgrades. Those are very different pictures of the same saved ratings.
Here is why I would be careful with the first interpretation. The fitted model explicitly gives different Elo baselines to troops, regiments, and hordes. When comparing a regiment to a horde, it subtracts about 204 Elo for the average formation-size effect before calculating the upgrade return. A troop-to-regiment comparison similarly removes about 108 Elo. In fact, I tested this directly: add 600 Elo to every horde and refit the model. Every horde’s implied points and PVR remain unchanged, because the horde-size adjustment absorbs the entire increase.
That is a useful diagnostic. The adjusted measure is good at asking whether any particular horde is efficient compared with the fitted horde baseline. It cannot answer whether hordes as a class are underpriced or overcosted. The near-1:1 returns are partly a consequence of how we set up the comparison (i.e. it’s partly tautological). The old and new snapshots also differ in rating runs, corrected profile handling, and baseline populations. Their contrast doesn’t isolate the effect of replacing Elo per point with PVR.
Actual units show how much variation hides in the median
Three regiment-to-horde upgrades illustrate the spread. All three are Core units in both sizes.
| Faction and unit | Regiment → Horde cost | PVR, Regiment → Horde | Adjusted upgrade return |
|---|---|---|---|
| Undead — Revenants | 135 → 215 | 1.10 → 0.95 | 0.69 |
| Orcs — Greatax | 130 → 230 | 1.22 → 1.12 | 1.01 |
| Kingdoms of Men — Halberdiers | 90 → 160 | 1.02 → 1.20 | 1.42 |
Under this particular model, Revenants lose ground when upgraded, Greatax are almost exactly on the fitted curve, and Halberdiers gain substantially. Those are useful leads for list-building questions. They aren’t direct estimates of how often one formation beats another on the table, and the ratings are still moving.
The pattern also differs by unit type when upgrading troops to regiments:
| Unit type | Matched families | Median adjusted return | Above 1.00 |
|---|---|---|---|
| Heavy Infantry | 23 | 1.18 | 16 |
| Infantry | 90 | 1.11 | 61 |
| Monstrous Infantry | 8 | 1.06 | 6 |
| Large Infantry | 43 | 1.01 | 22 |
| Cavalry | 35 | 0.96 | 9 |
| Large Cavalry | 16 | 0.79 | 0 |
The large-cavalry result caught my eye. None of those 16 troop-to-regiment upgrades beats the size-adjusted expectation. That’s worth investigating, particularly given how much cavalry depends on getting the right engagement. But the table cannot tell us whether price, durability, charge mechanics, or the mix of units explains the pattern. I wouldn’t use it alone to decide which size to field.
The simulator has a type–and it likes things that win duels
The broader results are consistent with something I’ve seen in earlier army-efficiency rankings. The model is fond of tough units that hit hard. Median PVR is 1.13 for Monstrous Infantry (16 profiles) and 1.07 for both Heavy Infantry (62) and Large Infantry (115). Cavalry sits at 0.93 (75).
The shooting split is even clearer. The 192 profiles with a parsed ranged weapon have median PVR 0.89, against 1.05 for the other 533 profiles. That group includes hybrid shooters, and the other group isn’t necessarily pure melee. This is a descriptive split, not a controlled estimate of shooting’s value.
The simulator includes movement, shooting, and most special rules within its duels. It cannot fully price the value of threat projection, deployment, screening, chaff, support, and scenario play across a whole army. A unit can earn its points without winning a fair one-on-one fight. I explored that same distinction in Is Shooting Dead in Kings of War 4th Edition?. The 4E army categories show a similar pattern. Specialists and Support each have a median PVR of 1.04; Core is 1.02, Auxiliary 0.96, and Warlords 0.90 (note that this isn’t considering the value of inspiring, auras, or command orders). The fitted baseline adjusts for price and size, not category or battlefield job. These numbers don’t establish that Auxiliary or Warlord slots are bad investments. They tell us what those profiles accomplish in the simulated duel environment.
Here’s a useful example of the difference between strength and value:
| Profile | Cost | Duel Elo | Implied points | PVR |
|---|---|---|---|---|
| Order of the Abyssal Hunt, Basileans Horde | 360 | 2,605 | 309 | 0.86 |
| Coral Giant, Trident Realm Titan | 195 | 2,020 | 266 | 1.36 |
The massive Order of the Abyssal Hunt is the highest-rated duel profile in this snapshot, but its price is high even by the fitted horde standard. The Coral Giant has a much lower raw Elo, yet comes in well above the model’s expectation for a 195-point Titan. That makes the Coral Giant the better relative-value example here. It doesn’t mean a Coral Giant beats the Order of the Abyssal Hunt horde in a duel.
What about factions?
Trident Realm has the highest median PVR in the current snapshot at 1.10 (33 profiles), followed by Ogres and Abyssal Dwarfs at 1.06 (47 and 34). At the other end are Nightstalkers 0.90 (24), Halflings 0.92 (35), and Ratkin 0.93 (27).
I would not turn that into a faction tier list. Twenty-three of the 24 simulated Nightstalker profiles have Mindthirst, which this simulator does not model (because inspiring is typically from other nearby leaders, not the unit itself). That’s a selected list of missing behavior, not a complete accounting of every rule or interaction the model may miss. We’ve also seen how easy it is for a unit-level efficiency ranking to miss what makes an army work. Ratkin were near the bottom of an earlier model while performing much better in competitive play. The difference between a good duel unit and a good tournament army is substantial.
Where I land
I still think these simulations are useful. They identify profiles that appear strong or weak for their cost in a controlled fight. They point us toward questions worth testing, such as why large-cavalry upgrades lag the baseline or why the same unit family scales differently across sizes. For the horde question, though, I don’t think we’ve established a universal tax. Nor have we shown that formation pricing is fair. We have shown that Elo per point can give a misleading picture of larger size units, while a size-adjusted value model answers a narrower question.
For now, the horde tax is a better question than it was a headline.
