On 28 February 2026 the Gulf states closed their airspace. A network carrier MarSci works with lost every departure from its hub overnight. By the end of the first week of March its bookings were down 81%. Not traffic but actual bookings. Forward demand had stopped and the entire industry was disrupted.
The question from the airline’s marketing team was simple: what should happen to the media budget, and how much of it is actually needed?
The airline is anonymised. It is a large network carrier with one Gulf hub, nine sales markets, around $8bn in annual revenue and $211m a year in paid media across eleven channels.
This is how MarSci approached it and the what the model showed.
Step one: a dataset that knows what happened
In similar econometric projects, data is the most important and crucial component. The industry often uses the saying: ‘Garbage in, garbage out.’ However, in our experience, data is never perfect. Some datasets sit with ground operations, others with branding, and others with the marketing agency. Getting access to all of them can be difficult, even for internal stakeholders. That’s why at MarSci we have created a different process. Each dataset required for the model is assigned a name and a stakeholder who is responsible for it. We then contact them directly, and it is our responsibility to obtain a single file from each of them. We then combine and clean the data. Although this is a challenging task for us, we have found that taking on this responsibility moves the process along faster.
MarSci built a single weekly file from January 2024 to July 2026 across nine markets. Metrics included: net bookings, revenue, marketing spend, impressions (for TV, OOH, paid search, Meta, YouTube, programmatic, metasearch, influencer, affiliate, sponsorship), trade co-op, CRM sends. But we didn’t stop there. We requested fares prices, how much of the schedule actually flew each week and finally we included third-party datasets.
Third party datasets are very important for any kind of an econometric model. Since the duration of the model was heavily disrupted, we wanted to keep our third-party data selection unbiased. Therefore we added: The Dubai floods in April 2024, the twelve day war in June 2025, when Iranian missiles hit Al Udeid and four countries shut their airspace overnight. The first flights out on 2 March as an event. We modeled the ceasefire on 15 April, the UAE lifting all restrictions on 3 May, the June flare up. Each one of those moments affect revenue and bookings more than marketing activity. Therefore, prior knowledge and information is extremely important for the model.
Step two: the control variables carry the shock
Without proper control variables, an econometric model could conclude that bookings collapsed in March because TV was switched off. TV was switched off because there were no seats to sell. The model has to be told that, in a variable it can use.
The following control variables went into Meridian:
Ops index. The share of normal capacity flown each week. A very important variable in the model since it carries the crash and the recovery so the media does not have to get credit that doesn’t deserve.
Fare index. Jet fuel went from $92 to $215 a barrel in March. As it is normal fares followed. Price tickets would follow a very specific pattern but the interesting detail is that the barrel affects all competitors and therefore prices would fluctuate for all airlines.
Ramadan, Eid and Christmas week. Seasonality impacts bookings, there is no surprise there. But how does the region answer to the seasonal moments such as ramadan?
Search demand index. The great thing about travel is that search demand is widely available within the organization.
For the purpose of this model, MarSci utilized an open source software by Google (Meridian). Knots is a hyperparameter representing seasonality. The ideal knot setting though requires testing and deep understanding of how Meridian works. Behind a simple decision of a number is hidden different MMM results.
Another challenging point – especially when using meridian is the similar patters in media spend. If TV, YouTube and OOH always burst together and always get cut together, no model can separate them. The first pass had exactly that problem and the chains did not converge. Once the plan reflected each channel’s real flighting, every parameter converged with R-hat below 1.02.
The model explained 93% of weekly revenue with a 2.6% error. A model without the knowledge and understanding of meridian could give even higher RSquare but the business value would be zero since it didn’t take into consideration the context of the market.
Step three: what the MMM cannot see on its own
Airline media is small next to revenue. Every channel is under 1% of turnover. Against normal weekly noise of 3 to 4% that is a faint signal, and a Bayesian model with default priors does the honest thing: it shrinks every channel toward the prior and reports that it is not sure. Paid search came out at an ROI of 2. The airline’s own lift tests put it closer to 4.
So the model was calibrated. Three channels had test evidence: paid search, metasearch and affiliate. Tight priors on those, everything else left open. That one step moved paid media from 4% of revenue to 7% and produced numbers the team could act on. Search 9.0. Metasearch 6.3. Affiliate 5.0. TV, Meta and YouTube at 1.2 to 1.4 with wide intervals.
The lesson is not that a channel does not work. It is that at this scale an MMM alone cannot see it, and brand channels should not be cut on model output without a test.
Step four: Where to actually invest marketing investment
Meridian ships with a budget optimiser. Run on the 2025 plan, it recommends keeping the $211m, moving money out of TV, YouTube and influencer and into search, metasearch, OOH and sponsorship, for 6.7% more incremental revenue.
The advertiser did not want to proceed with that recommendation. The optimiser only knows the response curves and two of those curves, OOH and sponsorship, reflected the prior rather than the data. Handing them 30% more budget on that basis is how money gets wasted and frankly why MMM don’t be trusted by companies.
This is where Causal Discovery Attribution comes in. CDA is a peer-reviewed model, published in the Journal of Marketing Analytics and open-sourced by MarSci, built directly into the MarSci engine.
What CDA does, in plain language
An MMM asks: how much did each channel contribute to bookings? CDA asks a different question: how do the channels actually relate to each other and to bookings? It reads the same weekly data not as a regression but as a map of cause and effect. From that map it can tell apart three things that look identical in a standard model:
- Channels that move bookings on their own.
- Channels that only move other channels.
- Channels that appear to “work” only because bookings moved first.
Why combining CDA with the MMM is unique
CDA doesn’t run in isolation it takes the MMM’s own coefficients as input. The MMM estimates the money value of each channel (its ROI); CDA supplies the causal structure (which channels genuinely drive outcomes, and through which paths). Optimising on either alone is dangerous: the MMM’s ROIs can be artefacts of priors or correlation, and a causal graph without ROIs can’t allocate a single dollar. Put together, every euro of recommended spend has to pass two tests the channel must be worth the money and there must be a verified causal path from that spend to bookings.
For this airline, the graph changed three things:
- Paid search and metasearch stayed as direct drivers. The MMM already trusted them.
- OOH and sponsorship showed no independent path to bookings in the data. Held, not increased.
- TV fed search volume with a one-to-two-week lag. Cut, not killed.
The numbers
With the CDA constraining the optimiser and a target ROI of 5 on every incremental dollar, the recommended plan came to $171m against $211m.
A 19% reduction in paid media, $40m a year, with modelled revenue held flat.
| Channel | 2025 spend | Recommended | Change |
|---|---|---|---|
| Paid search | $38m | $57m | +50% |
| Metasearch | $14m | $21m | +46% |
| Affiliate | $6m | $7m | +5% |
| Sponsorship | $22m | $19m | -13% |
| OOH | $19m | $12m | -38% |
| TV | $42m | $21m | -50% |
| Meta | $26m | $13m | -50% |
| YouTube | $16m | $8m | -50% |
| Programmatic | $10m | $5m | -50% |
| Influencer | $9m | $5m | -49% |
| Trade co-op | $9m | $4m | -51% |
| Total | $211m | $171m | -19% |
Forty million dollars matters to an airline that IATA expects to lose money this year.
What carries over to other businesses
Any business that had its demand switched off by something outside marketing faces the same problem. A war, a flood, a regulator, a pandemic. The model has to be told what happened or it will blame the ads or the data that you feed in.
The recipe does not change but the operational excellence support MarSci to achieve robust results within weeks. If you wish us to help you with your model, do not hesitate to reach out.