ROI attribution for non-traceable marketing channels
In a lot of cases, the impact of certain marketing channels is not directly traceable. For example in the case of TV advertising, we can not directly observe how a campaign influenced certain potential customers in their buying behaviour. In the case of TV, it is possible to measure the so-called direct impact of a spot, by looking on the website traffic of a product and flag users which visit it in a certain time window after the spot. This method became popular in the last years but has certain drawbacks. The main problem of it is, that only a small amount of the potential customers who watched the spot are visiting the website within the next minutes after the spot; nevertheless, all people who watched the spot are potentially influenced by it. As a result, most KPIs like for example CPO (cost per order) or CPV (cost per visit) look much higher than they are in reality – which leads to wrong conclusions. In most cases, it will look like as if TV is not a marketing channel in which it is worth to invest in – a demonstrable error.
Our task was to develop a model to attribute the correct impact of TV advertising (or other non-traceable marketing channels) only on the basis of data which is available in almost all cases: a time series of marketing activity in the channel (for example net costs spent per week over time) and a time series measuring an impact (for example orders of the product per week over time). This is a difficult task and initially, it was not clear if this is possible, since especially TV advertising has not a short but a long-term effect. Additionally, there are a lot of overlaying effects, like for example seasonalities and activity of other marketing channels.
It turned out, that apart from measuring the direct impact of a spot, it is also possible to estimate the impact of a campaign with an indirect methodology. The main advantage of this method is, that we are able to measure the whole impact of the campaign and not only the impact on a small subset of people visiting the website directly after the spot, which distorts the true KPI’s which are usually much better. Also, it gives us the possibility to build a prediction model which we can use to plan future campaigns better.
For measuring the indirect effect, we trained a prediction model which explains the time series of a KPI (e.g. orders per week over time) as an echo of our marketing activity in the channel (e.g. measured as net costs per week for TV advertising) plus the impact of other marketing channels, which we not necessarily need to know (but can improve our performance).
From this, we can derive an explanation for how the marketing channel affected our orders and can decompose our orders time series into a time series of orders attributable to our marketing channel and a time series of orders attributable to other effects. By this, we can give a plausible an accurate estimation of the true impact of the marketing channel. If we have enough data, the impact can be also measured over a long time range.
The methodology we use is related to Temporal Canonical Correlation Analysis, but uses an additional Bayesian model approach specifically developed for our problem to incorporate previous knowledge and make the solutions more stable. We already applied it successfully to different kind of marketing channels and products.
It turns out, that TV advertising is indeed a marketing channel which has a strong impact and for the first time, it was possible to measure the correct and long-term effect of TV advertising.