How Customer-segment MMM benefited this Client from the Financial Sector

How segment-level MMM showed a retail bank that its marketing budget increase was working, and that the acquisition drop was concentrated in two customer segments most exposed to external influences.

The challenge

For a retail bank running campaigns across a broad customer base, one blended number can hide more than it reveals. Acquisition fell 4.5% year over year. Marketing budget, meanwhile, rose 16% over the same period. That combination looked like a broken model, or a wasted increase. Aggregate marketing mix modeling could confirm the drop, and it could point to competition and external influences as plausible causes. However, it could not say which customers were actually driving the loss. Nor could it say whether the extra budget was doing anything at all. This case study describes how MASS Analytics rebuilt the model at the customer-segment level, and what that granularity revealed.

One acquisition number hid four different stories

A national or channel-level model would have reached for two obvious explanations for the 4.5% drop: rising competition, or a budget increase that simply was not working. Both explanations are visible at the aggregate level. Neither, however, is correct on its own. The bank served four structurally different customer segments: college students, private sector professionals, public sector professionals, and retired customers. A single blended number could not show that two of those segments were absorbing the entire loss. The other two, meanwhile, were stable or growing.

Four things the aggregate model could not see

• The bank assumed competitive pressure explained a large share of the acquisition drop. Once the model ran at the segment level, though, competition’s measured impact was far smaller than expected in every segment.
• A 16% budget increase should have offset most of the drop. Aggregated, it looked like the increase had failed. Segmented, however, it showed the increase was working, just outweighed by external influences that hit two specific segments hardest.
• College students and retired customers also lost some volume to external influences, but not enough to move total acquisition. The real damage, instead, was concentrated in the private and public sector professional segments, invisible at the blended level.
• The bank’s own online media was quietly doing double duty. It drove acquisition directly within one segment. At the same time, through a halo effect the aggregate model had no way to detect, it lifted a second segment too.

Each gap above only became visible once the model stopped averaging the four segments together. Table 1 summarizes what the segment-level rebuild measured and found.

Segment-level modeling separated direct media effects from halo effects between segments. Paid search, social, and other channels stayed in the model with standard treatment.

The solution

MASS Analytics rebuilt the acquisition model at the level that actually mattered: the customer segment, not the whole bank. Four segments, college students, private sector professionals, public sector professionals, and retired customers, were modeled individually. Each used a multiplicative specification, fitted to three years of weekly data. A nested paid search model sat alongside the segment models, too. It separated media-driven branded search from organic demand. And every segment model controlled for promotions, price, CRM activity, and market conditions.

Each segment kept its own specification

Each segment’s model used AdStock to capture the carried-over memory effect of advertising. It used diminishing returns to model channel saturation, and a weighted sum to combine channels run at the same time. Different variations of these transformations were tested and selected using genetic-algorithm model selection. As a result, each segment kept the specification that actually fit its own data, rather than one specification imposed on all four.

The nested structure is what found the halo effect

The nested paid search structure let the model separate a segment’s direct response to its own online media from any lift that activity produced elsewhere. That is what surfaced the halo effect between private sector professionals and public sector professionals. The four separate, unconnected segment views would not have shown that spillover on their own.

Three years of segment-level data made the comparison possible

Three years of weekly acquisition data, provided by the bank, was modeled separately for each of the four customer segments. None of it was pooled into one national series. Media spend and impressions were tracked at the channel level for every segment, too, alongside promotions, pricing, CRM activity, and broader market conditions. That let each segment’s model isolate media-driven acquisition from everything else moving at the same time. Consistent segment definitions across all three years were the real precondition for the comparison. Without a stable segment structure, the year-over-year swings inside private sector professionals and public sector professionals would have been impossible to isolate from noise.

Results and Impact

The segment-level rebuild produced four findings that reframed the acquisition problem entirely. Chief among them: TV, not any digital channel, was the single largest driver of new accounts across every segment. It contributed about 5,000 new accounts on its own.

TV anchored acquisition, but segments moved separately

TV was the single highest contributor to new accounts across every segment, delivering about 5,000 new accounts. It also showed the strongest synergy with paid search of any channel pairing in the model. That finding held at the aggregate level. However, the aggregate level could not show which segments were actually driving the drop. Private sector professionals and public sector professionals accounted for nearly all of the 4.5% decline. Both were also the most exposed to external influences, with acquisition losses of 10% and 17% respectively. College students and retired customers, meanwhile, absorbed some of that impact too, but not enough to move their totals. In fact, the retired segment grew 6% over the period, driven mostly by Radio and TV. College students, for their part, responded positively to YouTube run alongside paid search. A blended model would have averaged both findings away entirely.

A halo effect connected two segments the aggregate view kept separate

The private sector professionals segment’s own online media drove a 6% acquisition lift within that segment. A nested version of the same model went further. It showed that the same activity also produced a 3% acquisition lift in the public sector professionals segment, a spillover invisible to either segment’s model alone. On the strength of that finding, the bank was advised to make four moves. First, increase Social spend, the highest-ROI channel for both hard-hit segments. Second, add YouTube targeting private sector professionals. Third, increase Google Display, which carries the lowest CPA and the highest ROI. Finally, build segment-specific activity for public sector professionals, given its outsized contribution.

Related case studies

• Nested Modeling in MMM: how a major telecom operator used the same nested-modeling technique to map its full customer journey and found a 12% digital sales uplift.
• Kellogg’s: Weeks become days.: how automating the full MMM cycle cut an 8 to 16 week turnaround to about 7 days.
• Promotional Performance: isolating the true incremental impact of individual promotional mechanics for a retailer running complex, overlapping promotional calendars.