Abstract
This article systematically reviews 22 merger simulation models cited in 16 European Commission merger decisions between 2000 and 2018. It first analyses disagreements between the Commission and the merging parties over simulation design, finding increased disagreement over time. It then tests common beliefs about their price predictions. Econometrically estimated models generally predict lower price increases than calibrated models, while merging parties do not systematically produce lower estimates than the Commission, nor are telecoms mergers treated more harshly. Where both sides use econometric models, estimates tend to converge. We argue that this consistency is a promising basis for restoring confidence in simulations.
I. Introduction
Merger simulation has been something of an unfashionable tool lately in Brussels. The Commission first relied on one in Volvo/Scania in 2000.1 Use peaked in 2014–2016, when nearly half of all the merger reviews that included a simulation took place, but the last published decision reporting simulation results dates from 2018.2,3 Simulations may still be used internally, as screens or as support for structuring a case, but their appearance in published decisions has become rare.
Commentators often explain this decline by the adversarial nature of simulations, where they are either put forward by the Commission (where models are said to be built to find harm) or by the parties (whose models are said to be built to find none). If the simulations are so susceptible to what each side would like to show, then it is reasonable to question their usefulness.
The main contribution of this paper is to test this conjecture. This is made possible because Commission decisions are transparent about how simulations are built and what they predict. Our conclusion is not that simulations are infallible. It is that the real fault line runs between methods, not between sides, and that this distinction has been underappreciated.
The paper does not assess whether simulations accurately predicted what happened after mergers. It measures something narrower: whether adversaries using different models arrive at similar numbers.
II. Two ways to build a merger simulation
A merger simulation combines a model of how customers choose between products (demand) with a model of how firms set prices (supply). It then asks what happens to prices when two rivals are combined. A key distinction is how each model’s parameters are obtained.
A simulation is calibrated when its parameters are set from existing information rather than estimated: typically market shares, profit margins and diversion ratios, meaning the share of customers who would switch from one product to another if the first became more expensive. Calibrated simulations are cheaper and faster. The Commission’s calibrated simulations typically assumed linear demand. The price of that convenience is that more rests on assumptions chosen by the analyst.
A simulation is econometric when demand (and sometimes supply) parameters are statistically estimated from detailed data on sales and prices. These models can capture richer substitution patterns and can even inform market definition. They also demand more data, more time and more specialist judgement, and the most sophisticated versions (the so-called BLP and random-coefficients models) have been accused of producing unstable results.
Of the 22 models in our sample of simulations, 12 were calibrated and 10 estimated econometrically. Both model types appear in the Commission’s models as well as the merging parties’, and the paper treats the choice between them as one of the main outcomes of interest.
III. What the two sides actually argue about
We begin with a qualitative review of how the decisions describe each simulation, classifying every modelling feature as uncontested, debated (the other side objected and the author responded) or criticised (the objection was left unanswered). The exercise is unavoidably somewhat subjective, but several patterns stand out.
Firstly, we find that models travel, that is, many simulations build on a model used in an earlier case. The Commission’s Unilever/Sara Lee4 model closely followed the merging parties’ model in Kraft/Cadbury,5 and in some cases reuse was welcomed. In the later telecoms cases, however, the merging parties complained that the Commission kept recycling modelling choices without asking whether they could be improved.
Reusing previous models, however, did little to settle the more fundamental questions. The sharpest disagreements concerned model choice and the realism of the assumptions. In Oracle/PeopleSoft6 the dispute was over which type of auction best described the market. In Outokumpu/Inoxum7 it was over whether the model captured a realistic pre-merger world, and there and in INEOS/Solvay/JV8 the Commission rejected the parties’ proposed adjustments as resting on “extreme” assumptions. Countervailing factors such as buyer power, entry and product repositioning were contested in most cases, with the Commission tending to argue that these play out over the long run and are too complex for a merger review. Efficiencies and quality were similarly divisive. The Commission sometimes built the parties’ efficiency claims into its own models, but at other times it challenged their technical basis, or was itself criticised for ignoring likely quality improvements.
The inputs used in the simulations were argued over too, in particular the treatment of the “outside good” (customers who buy nothing in the market) and the data used to calculate diversion ratios. The sophistication of the models offered no protection from disagreement. The most complex econometric models, BLP and random coefficients, were the ones attacked as unstable or overly sensitive, whichever side had built them. In Hutchison 3G UK/Telefónica Ireland,9 the Commission ultimately placed no weight on its own BLP model after the parties made exactly this argument.
On the other hand, not everything was contentious. For example, the use of instrumental variables and the modelling of information asymmetries were rarely disputed. The trend over time is striking, however. When the cases featuring simulation are grouped by the DG Comp Chief Economist who oversaw them, it can be seen that the share of modelling issues that did not face any debate or criticism at all (green line) decreased over the period, whilst the share of issues that were debated (orange) or criticised (red) rose.
Figure 1 – Share of uncontested/debated/criticised modelling issues, by Chief Economist
Notes: For each Chief Economist, each modelling issue across all the merger simulation models that were put forward in their tenure is classified as uncontested, debated or criticised, and the share calculated for each. The numbers in parentheses relate to the number of cases featuring at least one merger simulation model that each Chief Economist oversaw.
IV. Do the Commission’s models always find more harm?
Our data set covers 22 merger simulations produced in 16 Commission merger cases, yielding 780 predicted price changes: 272 baseline estimates and 508 sensitivity estimates. The raw data seem at first to confirm the conventional wisdom. The Commission’s simulations predict somewhat higher price increases than the parties’, although the difference falls just short of conventional statistical significance. These averages, however, hide who is using which method. We therefore turn to regressions that control for market concentration, the change in concentration caused by the merger, whether efficiencies were modelled, the procedural phase and other factors. Two results stand out.
The first is that method matters, but not uniformly. Among the Commission’s own simulations, econometric models predict lower price increases than calibrated ones, by about 2.5 percentage points in telecoms and about 11.7 points in other industries. For the merging parties’ simulations the pattern reverses, with econometric models giving the higher predictions. This comparison should be read with caution, as it rests on very few models. In telecoms it effectively sets the parties’ random-coefficients model in Hutchison 3G Italy/Wind/JV10 against the calibrated model submitted in Hutchison 3G UK/Telefónica Ireland.11 The first predicted large nominal increases, which the parties argued would be outweighed by quality gains; the second predicted “negligible” increases.
The second result is that the Commission does not always find more harm. When both sides use econometric models, the parties’ predictions are higher than the Commission’s, by about 2.4 points in telecoms and 9.5 points elsewhere. When both use calibrated models, the Commission’s predictions are higher, by roughly 10 to 12 points. So the modelling approach affects not only which side predicts more harm but also how far apart the two sides are, and the gap is considerably smaller under econometric methods.
Why should calibration produce wider disagreement? Our intuition is that calibrated models leave more room for assumptions, and therefore for divergence. The CMA’s 2024 review of Vodafone/Three12 offers an illustration. Its econometric baseline implied consumer harm of around £216 million a year. Sensitivities using calibrated demand parameters instead produced estimates of up to £900 million. The CMA itself described calibration as more approximate and placed limited weight on the specific harm estimates it produced.
There is another possible reading. The Commission typically assumes linear demand in its calibrated models and treats the results as a lower bound. More generally, where its predictions were lower, it may have chosen more conservative assumptions, to show that harm remains even on terms favourable to the parties.
V. Is the telecoms industry special?
Critics argue that telecoms mergers (often “4-to-3” deals) are an easy target for finding high predicted price rises in a simulation. Fixed costs are high, customers switch between a few operators, and so models mechanically yield high margins, high diversions between the merging parties and large predicted price rises. If so, simulations might have flourished there because they help the Commission build a case.
The data do not support a simple version of this claim. Simulations in telecoms do predict higher price increases, but only for econometric models run by the Commission. For econometric models run by the parties, and for the Commission’s calibrated models, the telecoms-versus-other-industries difference is either statistically indistinguishable from zero or, in some specifications, runs the other way. For the parties’ calibrated models, predictions are actually higher outside telecoms.
One explanation for why simulations are clustered in telecoms is that early cases created precedents that later cases built on. And in an industry with bundling, nonlinear tariffs, price discrimination and constant promotions, market shares and simple price comparisons are poor predictors, so the case for a richer model is stronger.
VI. Is convergence of outcome just a sign of agreement?
One might argue that consistency between two interested parties is a weak test of quality. They could converge on the wrong answer, or the Commission’s choice of conservative assumptions could mask its true view. Moreover, the paper’s own approaches in its qualitative and quantitative analyses carry several limitations.
First, our qualitative classification of what was “contested” is subjective and limited to what appears in published decisions.
Second, in the quantitative analysis, the sample is small. There are 22 models, and several comparisons rest on only a handful. The results are also fragile in places. When standard errors are clustered by case rather than by model, the econometric-versus-calibrated gap for the Commission’s telecoms simulations loses statistical significance. Other robustness checks, such as weighting by market share, adjusting for quality gains, and changing the controls, leave the broad picture intact but shift individual coefficients.
Third, redactions mean that some predictions could not be included at all. This could mask a selection problem, whereby only certain types of (and not a random sample of) simulations are published.
Fourth, and most important, the paper does not test whether simulations predicted actual post-merger prices. Outside the EU, that literature is mixed. Some studies find simulations predicted outcomes well in cereals,13 beer,14 hospitals15 and painkillers.16 Others find weak performance in airlines17 and motor oil and syrup.18
Nevertheless, we posit that if agreement between adversaries is a reasonable proxy for robustness, then econometric models perform well on that measure. We also argue that the complaint “parties can game simulations” applies at least as strongly to simpler price-pressure tools such as GUPPI,19 which rest on more structural assumptions, yet are less criticised. If gaming is the worry, that argument seems to favour fully fledged simulations rather than weakening them.
VII. What this means for merger review
The most immediate implication is for how disagreements over simulations are handled. Given that a dispute over a simulation is usually a dispute over its inputs, it is better settled by examining those inputs than by discounting the exercise. Econometric models, under which predictions converge, are the natural starting point, and telecoms simulations provide a useful template. At the same time, we would argue that simulations should not be used on their own. They belong alongside price-pressure methods such as UPP and GUPPI. As a next step, a useful exercise would be to compare simulation predictions with GUPPI, to test whether a method resting on more structural assumptions leaves more room for disagreement, and similarly to test divergence between the Commission and the parties when price-pressure methods are used.
Moreover, we believe that the case for using simulations in merger review is getting stronger. More mergers now turn on dynamic competition, quality, efficiencies and two-sided platforms, where market shares and simple price comparisons say little. In Hutchison 3G Italy/Wind/JV, T-Mobile NL/Tele2 NL20 and Vodafone/Three,21 the parties submitted novel simulations that put efficiencies and quality directly into the model. This should be encouraged. At the same time, platform mergers in the EU have not seen a simulation deployed, despite relevant academic literature22 (see, for example, cloud-gaming concerns in Microsoft/Activision Blizzard23 as the kind of case where an improved simulation might help).
Time pressure explains only part of the decline in the use of simulations. Deadlines are tight, and both the Commission and the CMA have said they limit what can be modelled reliably.24 But Phase 1 and Phase 2 reviews have not become shorter, so deadlines alone cannot explain why simulations disappeared from published decisions after 2018. Earlier investment in a simulation setup during pre-notification would help, as would newer methods that cut the time needed to estimate structural parameters.
VIII. Conclusion
The evidence from the Commission’s decisions points away from the worry that simulations simply reflect the interests of whoever builds them. What seems to drive the differences is how the model is constructed. When demand is estimated from data, the two sides’ predictions sit reasonably close together. When it rests on calibrated assumptions, they drift apart. We do not test whether simulations got actual post-merger prices right, and that question remains open. But it appears that, more often than not, it is the input used for the simulation that is most contested between the two sides. As such, simulations deserve a larger place in EU merger control than they have had since 2018. That is all the truer as mergers look less and less like the cases that existing shortcuts, such as price-pressure methods, were designed for.