Setting The Communications Budget

Setting The Communications Budget Paul Dyson Chairman, D2D Limited Introduction Setting the communications budget is a topical question - and so it should be. Since advertising for many brands is the largest chunk of the marketing budget it is vital to get it right. A 5% or a 10% improvement in efficiency can mean a significant saving. This paper discusses how we can help answer the communications budget question (and linked questions such as “How do I divide it across my portfolio of brands?”). The key is not a new statistical technique or model, but a modern, contemporary approach to statistical analysis, tailored to the specific problem and incorporating quantitative and qualitative inputs. This analysis needs to build itself around the inputs available – via tailored software – as opposed to the traditional approach of shoehorning inputs into a statistical analysis package. And the analysis needs to pull together knowledge and learning from not only the data but also the marketing team - and its suppliers – to produce a realistic and useable model. Partly this is good statistical practice, but the main reason for this approach is that measuring the effect of advertising is probably the most difficult marketing nut to crack in modelling terms. The Right Modelling Approach It has not been unknown, for modelling agencies to take clients’ sales data, disappear for 8 weeks and return with a model. In the 1980s and early 1990s this was perhaps acceptable. Brand managers were still learning about their brands – they were quantifying elasticities for the first time, understanding how advertising worked, and putting a figure on the amount of “pantry-loading” their promotions were producing. Nowadays however, the marketing team is more clued-up and an outside modelling team can learn from them. Brand managers tend to have a good idea of their price elasticity; know whether promotions cause increases in penetration or weight of purchase; and have a good fix on the value of increasing distribution by an extra 1%. Any attempt at modelling sales data to help find the optimum advertising budget needs to tap into this knowledge and other learning within the corporation. Even if the focus of a model is advertising, it should include factors for pricing, promotions and distribution, so it is vital the model is consistent with previous learning (or provides convincing evidence to the contrary). This points to the need for frequent interaction and communication between the modelling and marketing teams:  Pre-modelling interviews with key personnel (brand/marketing team, finance, sales, advertising, media, etc.) to help understand what drives the brands being analysed, and what different departments want from the project.  Frequent working sessions during the modelling process to ensure models are realistic (and to help get buy-in).  Availability of the modelling team beyond the final presentation to help integrate the model into planning and ensure its correct use. Also to feedback learning as the model is used. The modeller must be an all-round consultant, with good statistical skills, good communication skills and a good marketing/business brain. A number of other factors have changed over the last 10 years, which affect the way statistical analysis should be carried out: 1. Data have improved. More data, more frequently, from a variety of sources. This is part of the reason why marketing teams are now much more knowledgeable, but there is now sometimes too much data to handle easily. Today’s analysis has to work with different types of data (sales, research, qualitative) based on different time periods and samples. 2. Technology has improved. On the negative side, software advances allow us to produce ever more data on subgroups, target audiences, etc. Sometimes people forget that they are basing decisions on very small, unreliable samples. It is also easy to over-trust the output from software without questioning the results. On the positive side, we have the technological power to handle today’s mass of data, and to develop software to ensure that our analysis is tailored to the project. 3. A movement from brand-centric to corporate-centric decisions. Many companies have grown globally and have a portfolio of international brands. Marketing issues involve allocation of resources between brands, or between countries for a given brand, or both. If modelling one brand has become complex, modelling a portfolio across several countries is even more so. But it can and should be done. Significant efficiencies can be gained by trading-off the elements of a portfolio. So how should statistical analysis have evolved to cope with today’s demands? The process should enable a free, unhindered flow of information between the modelling team and the client. The project should be viewed as team work – the modeller needs access to all necessary data and the marketing team needs to understand, in their language, what is in the model and why. But most important is that the modeller should be flexible enough to allow his analysis to evolve around the available data. For instance, suppose the marketing team has weekly advertising data and monthly sales data. The normal solution would aggregate the advertising data into months to match sales, thereby losing some information (two months may have the same totals, but different weekly patterns). This is necessary because most statistical analysis packages only accept data with consistent time periods. A much better approach, however, is to develop bespoke software to handle a two-part model, which works simultaneously on both weekly and monthly inputs. Another example: say a brand has always advertised and promoted at the same time, to the extent that it is impossible to distinguish their relative contributions using normal statistical methods. There are various opinions amongst key people about the contribution of each factor to sales. These opinions can be investigated and incorporated into the model, since they represent our best guess in the absence of other information (which is better than ignoring the issue). This also gives realism to the model and helps it gain acceptance amongst possible users. Further, the team is aware of the gap in information and steps can be taken to quantify it in the future (e.g. by regional tests). In truth, there are times when the rigidity of statistical theory hinders analysis. Significance tests and r-squareds are important, but there is a point when common sense tells us to override theory and adopt a pragmatic solution. Bespoke software gives us the freedom to link rolling weekly tracking data to discrete monthly sales data; to incorporate qualitative opinions; to investigate and quantify difficult concepts such as price points (where the relationship between price and sales “jumps” at certain points - e.g. when price increases from 99p to £1). However, although this might be seen as “less statistical”, doing it properly requires experienced statisticians. If the theory is “loosened” somewhat, it is vital that the modellers know what they are doing. Getting a model that fits is the easiest part. Building one that makes sense and works going forward is the difficult bit. Understanding the impact of different inputs and what is achievable (and reasonable) with a model is vital. An experienced practitioner is a must at all stages. Further, this approach is probably most relevant when trying to model advertising - of all marketing inputs the most subjective and hardest to measure. To summarise, today’s analysis must reflect today’s information and needs. Attempts must be made to pull together all relevant sources of information, from hard sales data, through consumer attitudes to qualitative opinions about how the brand works. Bespoke software should be written (if necessary), to handle the analysis and to present the (usually complex) results in a straightforward, usable way. Finally, and most importantly, the process should not stop once the models have been delivered – they should be used, monitored and improved over time. Setting Communications Budgets. So how do we apply these principles to setting budgets? The starting point must be an analysis of how advertising has worked in the past, - but advertising is perhaps the hardest marketing lever to quantify. Advertising is often a small component of sales, compared to factors such as price and promotions; the effect is spread out over time, since people do not necessarily respond immediately; and it can have both short and long-term effects. All these factors make it difficult to tease out advertising effects using econometric modelling alone. There are however, some key pointers when modelling advertising:  Minimise aggregation – our experience suggests weekly data improve the ability to detect advertising effects from around 60% to 90%. Alternatively, looking at regions rather than national data might prove fruitful, since it can introduce more variation into the data.  Use other information, such as tracking studies. Advertising awareness data is engineered to highlight the effectiveness of advertising, so will be a good pointer to whether we might find advertising effects on sales  Ask key people their opinions – did the brand team think the advertising worked – or was it simply that the sales team could get more shelf space because a campaign was coming up?  Use database averages/published results – what is the % return for similar products; what is the expected relationship between short and long-term effects? All these point to a pragmatic approach. Working with different data sources on different bases; including qualitative opinions; using databases to channel the models in the right direction. For example, one project we worked on showed strong advertising effects all year except for the January sale. At this time sales dropped anyway (as they do for many brands post-Christmas) and the standard model was confused into believing there was a negative advertising effect. The client was unique and it was difficult to estimate seasonality from other competitors, who didn’t advertise. So we looked at advertising effects at other times of year and used this to estimate the probable effect in January into the model – a qualitative estimate, but a more logical solution that was agreed across the whole team. Whatever the modelling approach, once a satisfactory advertising effect has been identified, the key element for budget-setting is to consider the return each period against the weight of advertising in that period (the response curve). This allows us to start to understand how the return varies with weight. UK FMCG Brand Sales generated by advertising Monthly Advertising Investment Chart 1 – Advertising response curve In chart 1 each diamond is a month, showing the actual advertising investment plotted against the estimated return. Generally, as we increase investment we get a bigger effect and the line on the chart shows the relationship our model has estimated. It is non-linear: as we increase advertising weight per week (the horizontal axis) advertising effects start to level off. Diminishing returns set in. It is vital that we identify this if we are going to use modelling of this sort as a budget-setting tool. In theory, we want to be somewhere in the circled area – enough weight to take advantage of the steep increase in returns to the left of the curve, but not too much so we move into the inefficient flat part to the right. So what is the optimum monthly level and hence the optimum budget? Once we have identified our response curve we can start to make recommendations. The curve shows expected sales generated from advertising for different levels of investment. If necessary we can convert this to profit; and we can add on the cost of advertising as a straight line (chart 2): UK FMCG PRODUCT - TV Advertising 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 0 100 200 300 TV GRPs Sales from Advertising (£) Profit from sales Cost of advertising Sales (£m) 400 500 600 Chart 2 – Profit and cost of advertising We can then subtract cost of advertising from profit from advertising to get a net profit, which in this case has a maximum at 130 GRPs per month. Alternatively, if we wanted advertising to generate more sales (if sales were the priority over profit) we could advertise anywhere up to 300 GRPs per month and still make a profit from advertising (chart 3): UK FMCG PRODUCT - TV Advertising 1.5 1.0 0.5 0.0 0 -0.5 TV GRPs Sales from Advertising (£) Net profit Max profit at 130 GRPs per month Sales (£m) 100 200 300 400 500 600 Chart 3 – Optimum monthly weight. Once we have identified the optimum budget level, we can use the model to assess the likely improvement in campaign effectiveness. In this case, 130 GRPs per month is less than normal, allowing us more weeks on air, and although we do not quite reach the peaks achieved with our usual plan, the extra weeks produce about 15% extra sales for the same budget (chart 4). UK FMCG Product - Optimised Phasing 1350 1200 1050 +£1m = 15% GRPs/Sales 900 750 600 450 300 150 0 Jan-95 Jul-95 Jan-96 Jul-96 GRPs Sales (000s) Jan-97 Jul-97 Jan-98 Jul-98 Optimum GRPs Optimum Sales (000s) Chart 4 – Predicted sales from advertising under different media plans This is a straightforward example but shows the process for identifying optimum budgets. It is not difficult to extend the analysis to include different-shaped response curves; seasonality (e.g. more sales per GRP in seasonal periods); cost variations; and so on. With one brand it is relatively straightforward to find the optimum budget, but it becomes more complex when we consider allocation:    across a range of media channels across countries for international brands between brands (and/or variants) in a portfolio (allowing for halo or cannibalization effects). This is because an optimum budget determined for a brand in isolation may not be optimum when account is taken of other brands or countries. It may be more profitable to take some budget away from one brand and give it to a more profitable sister brand. Thus it makes no sense to work on each brand in the portfolio in isolation – it is vital to consider the whole portfolio (countries, brands or media channels) and optimise at corporate level. To give a taste of what is possible here is a (heavily disguised) case history. In this, we looked at relationship between three variants in a brand family – a parent brand and two sub-brands. Table 1 shows how initial modelling suggested that the variants’ advertising was most successful in generating variant sales - at twice the level of the parent advertising: Advertised Brand Parent Parent Sales per Variant 1 GRP Variant 2 £8,000 £16,000 £16,000 Variant 1 Variant 2 Table 1 – Advertising ROI for own brand However this ignored the interaction between the brands and their advertising – heightened here by the fact that they all used the same brand name in some form. If we add in interaction effects, we find that many of the sales generated by variant advertising came from parent brand sales, whereas parent advertising generated mostly new sales from competitors (table 2). Overall, parent brand advertising was more effective at generating increased total sales: Advertised Brand Parent Parent Sales per Variant 1 GRP Variant 2 £8,000 £3,000 £4,000 £15,000 Variant 1 Variant 2 -£11,000 -£8,000 £16,000 £4,000 £9,000 £3,000 £16,000 £11,000 Table 2 – Advertising ROI analysis including halo and cannibalization effects The next stage is to develop software to identify the optimum budget allocation between the three brands. Chart 5 shows a screen shot where each of the response curves have been input into software developed specifically for this client. Some of the curves are upside down, since they represent cannibalization by a variant from the parent brand: Chart 5 – Response curves in bespoke software Similarly, the curves along the diagonal show the steepest response, as we would expect since they represent the impact of advertising on the advertised brand itself. The software allows us to input a given budget and, using the response curves, works out the best way to allocate this across the three brands. It does this by finding the monthly advertising laydown for each brand that maximises total profit across all three. Chart 6 – Optimum allocation of a £10m budget For a £10m budget it allocates 51% to the parent and 21% and 28% to the variants (chart 6). Using the response curves it works out that this would generate profit from advertising of around £12m, giving a net profit from advertising of £2m. This is vastly different from the initial finding that variant advertising produces twice as many sales as parent advertising (table 1). On that basis we would have allocated more budget to the variants than the parent. Importantly, the software allows us to interact with the solutions. For instance, we might find it unacceptable to spend less than a quarter of the budget on any one brand, so we might set a lower limit of 25% for each. We could re-run the analysis and see what this does to the recommendation (and how far we move from the optimum). Chart 7 – Optimum monthly laydown by brand Chart 7 shows the individual laydown by variant for the optimum allocation of our £10m budget together with the predicted impact on sales. With the current scenario, the parent brand has virtually zero sales generated by advertising (its sales line is close to the zero horizontal axis). This is because the variants cannibalise parent brand sales when they advertise and the parent’s advertising generates enough new customers to counteract this. But how do we identify the optimum budget? In the example above we specified the budget. However, the software can step through a variety of different budgets, working out the optimum allocation across brands in each case and calculating total sales and profit from advertising aggregated across all the variants. Chart 8 shows how these elements change as the budget increases. Chart 8 – Sales, profit and net profit at different budget levels Chart 8 is very similar to chart 3, where we looked at an individual brand and its response curve. Now, however, we are looking at the aggregated impact for our portfolio. The bottom line shows the net profit from advertising –the profit from advertising-generated sales minus the cost of advertising – for different budgets. It peaks at a budget around £5m, but is fairly flat in the range £3m to £8m. This would maximise profit from advertising. If the aim were to maximise sales we would suggest a budget around £15m – where sales are maximised and advertising just generates enough profit to pay for itself. If our objective mixed both profit and sales targets, then we simply identify the relevant point on the curve. Summary The case history shows how we can use modelling and bespoke software to help key budget decisions. The example concerns budget allocation across a portfolio of brands, but the same approach can easily be adapted to a portfolio of countries or channels. Indeed, it can be multi-dimensional - we have applied this approach to a portfolio of brands across a range of countries and to a portfolio of brands across a range of media channels. The analysis is complex. It requires bespoke software to present the complexity in a useable way. This complexity means it is important to: 1. Treat the results with care. Use the software as the starting point for discussion, not the final answer. Test the recommendations out if they are very different from normal practice – e.g. by region or by degrees. 2. Use only experienced analysts to build the models/software, and ensure they are involved in using the results - it is easy to stretch a complex model beyond its limits. 3. Ideally, run the software as part of a working session - it is rare that the “optimum” is palatable in every sense (it may suggest no support for less profitable brands): a working session allows the software to be constrained. One might argue that advertising is such a difficult factor to measure, and there are so many imponderables that budget-setting will never be an accurate science. True! However, an approach like the one outlined – where there is a framework for capturing as much relevant knowledge and information as possible - allows us to get a little more scientific and accurate than we might otherwise be.