Category

Allgemein

Data quality cleaning using machine learning

By | Allgemein

Data quality cleaning using machine learning

New, more effective ways to avoid bad response effects, sampling bias and fluctuating KPIs

Some respondents click rapidly through the questionnaire, some erratically. The random sampling in a survey is often a far cry from a chance selection. Excessive quota systems increase the problem instead of solving it. Internal customers often ask themselves why an NPS (or other KPIs) has decreased or increased and can find no reasonable answer. Misgivings regarding the data quality can destroy the trust in an entire study. Luckily, there are new, effective ways to counter this.

The bad response effects example

Many market researchers have developed their own “practical” rules to counteract this. Overly short survey times or too many affirmations lead to the exclusion of a respondent. Occasionally, test questions are introduced to check whether a respondent is “paying attention”.

The problem is: what exactly is the right survey time that should be used to exclude respondents? Which rate of agreement can be used to differentiate a fan from a dishonest respondent? What exactly does the test question say about the survey behavior of the person?

The answer to this question is not known and can hardly be answered from data using typical means.

Self-learning analysis methods, also known by buzzword “machine learning,” offer the chance to learn of complex relationships between indicators and their effects based on data.

Bad response indicators include, for instance, the length of the interview, the response to a trick/consistency question, the number of identical responses in a row, and many more. The trick no longer tries to assume the meaning (= influence) of these indicators for the results (arbitrarily) but rather uses the data to discern the causal influence. This is possible with cause-effect analyses based on machine learning (especially NEUSREL). The analysis yields the following results. The graph shows that customer satisfaction is extremely high for interview lengths under a certain number of minutes (the scale is normalized to 100 here).

diagrammsatisfaction_2

This makes it possible for us to derive reasonable decision rules. Below a time of 20 (normalized) the satisfaction is obviously massively biased. On the right side, you will find a scatter plot containing the same variables, and you see a very different relationship there. Why? It could be the case that the fast respondents are younger people who are per se less satisfied with the provider. Machine learning delivers causal, corrected insights instead of the typical spurious correlations.

If the data is used for a driver analysis, it is recommended to not exclude the suspicious cases at all. Because the bad responses are described well by the indicator variables, a driver analysis based on machine learning (especially NEUSREL) simply filters out the bad response effects. The sorting out of cases would introduce a new but unknown selection bias to data.

The sampling bias example

Distortions in random sampling can occur, for example, as a result of a willingness to participate among the subjects that does not differ randomly. By measuring indicators of this willingness to participate, such as age, sex, income or a trait such as the “need to communicate,” machine-learning-based cause-effect analyses can determine the relation between indicator and response behavior. If there are too many seniors in your sample, for instance, you then know how this will distort each question and consequently you can correct this effect.

For the use of data in driver analyses, it is sufficient to make good indicators part of the driver model, because it automatically subtracts out all bias effects. However, this is possible only for driver analyses based on machine learning (especially NEUSREL). In this manner, an overly strong quota system, which itself creates new quota-related biases, can be avoided.

p.s. More on NEUSREL and Universal Structural Modeling here2

Lowering returns quotas: breakthrough via causal analysis.

By | Allgemein

How leading returns researcher Professor Walsh discovered effective levers to lower the returns quota via causal analysis.

The extremely high returns rates in e-commerce are an unsolved problem – a problem that represents a challenge to managers due to the tight margins in online retailing. Traditional BI analysis results very often turned out to be misleading in the past. The reason is known as “spurious correlation”.

Success Drivers  therefore identified, for example, that horizontal striped clothing was sent back less often. A classic spurious correlation: horizontal stripes are found particularly on cheap products, for which the returns rate is generally lower. Under otherwise similar circumstances, horizontal striped products are sent back MORE not less. This is a finding that can only be discovered via causal analysis.

The returns whisperer

For this very reason, Professor Walsh from the University of Jena has been involved in causal analysis with his team of researchers. His various studies provide many new insights into what drives product returns.

In a study, which appeared in the Harvard Business Manager, he and his co-invetigator found that the use of penalties and obstacles does not pay off in the long run. Their study showed a clear strategic path: return rates can be reduced by providing customers with relevant product-related information and making the buying process as hassle-free as possible. Virtual try-ons, avatars and specific product evaluations from peers are very helpful in this respect. The latter is a factor with very high ROI as the investment is low.

Self learning causal analytics covers up

With the current study the research team around Gianfranco Walsh were also able to identify a key driver of customer returns. Using consumer survey data and causal analysis, they found that the customer-related reputation of an online retailer alone has a significant effect on the return rates.

This reputation is determined by the following components: customer-orientation, being a good employer, being financially sound, perceived product quality and social responsibility. Although this measure may seem very abstract, the new causal model enables very precise measurement of the effect of bad press and to quantify the effects on the margin.

“There are also many company-specific returns levers. They may be identified most successfully via self-learning causal analysis – as enabled by the NEUSREL software”, Walsh recommends.

p.s. More on NEUSREL and Universal Structural Modeling here.

Conjoint Analysis: Price & Product Optimization Using Holistic Conjoint Analysis

By | Allgemein

Conjoint Analysis: Price & Product Optimization Using Holistic Conjoint Analysis

It is an open secret in the industry. Conjoint Analysis Software often delivers strange results forcing analysts to tweak datasets or analytic parameters so that results will have some face validly. Only this is not a procedure that produces correct results. The phenomenon is a clear indicator that something is still wrong with today’s conjoint analysis approaches.

The challenge with Conjoint Analysis

Many advances to Conjoint Analysis approaches have been developed. There is ACA (Advanced Conjoint Analysis enabling for partial feature profiles), CBC (Choice Based Conjoint switching from preference to choice) or CSC (Constant Sum Conjoint enabling for a distributed choice).

The basic principle is to conduct several choice experiments by manipulating causes (i.e. features). Due to practical reasons, it is not possible to assess all possible feature combinations at one respondent. This is where the issue evolves. That is why the choice of feature combinations per person is randomized. This would work if samples were large enough, but they rarely are in real life studies!

An example clarifies

Imagine there is a study where actual consumers of a food brand are more likely to choose the brand – simply because they have already tried it. But this feature “have already tried” is not part of the conjoint experiment. If features are randomized, there will be features or feature combinations with significantly higher ratios of persons who have already tried the product.

Since they are much more likely to buy, conjoint analysis will “think” those features or feature combinations must be important. The problem is not solvable by quotation nor by other measures in advance because normaly we don´t know every “disturbing factor” in advance.

Why a solution of the Conjoint Analysis dilemma would be so valuable

If we could consider all the biasing factors of a conjoint, market researchers would enjoy several advantages:

  • First and foremost, it would very much increase the trust and confidence in the validity of results. You can be much more confident that derived recommendations will work.
  • An analysis with all relevant factors will enable a holistic view of the decision process. It will not just focus on the choice decision but will also quantify the impact of product features and prices towards brand reputation, brand awareness, consideration, loyalty or recommendation.
  • Finally, such a study would be a single source for multiple information needs. It may serve for brand positioning, segmentation and customer journey questions.

The powerful solution: Holistic Conjoint Analysis

For a holistic conjoint analysis approach we need to collect several pieces of information:

  • Several choice tasks per respondent with randomized feature profiles.
  • Attitudes and image of brand.
  • Data about the outcomes of interest, such as purchase intention, willingness to inform, brand consideration
  • Data about the respondents such as demographics, their situation, and their history. Are they already customers? Are they bargain seekers? Are they familiar with the costs? All those items that might have an impact on their desired outcomes.

The ultimate question is ‘how do we identify which changes would improve the outcome?’ It is a challenge to conventional methods as they are built to fit parameters of a known model.

What is needed is a method that builds a prior unknown model in light of data, a model that incorporates indirect causal effects and does not neglect nonlinearity or interaction of those unknown relationships. This is exactly the strength of the Universal Structure Modeling approach that is performed by the NEUSRELTM Software.

The good news: Holistic Conjoint Analysis is proven, reliable, tried and tested

We compared dozens of studies on the validity of Holistic vs. Conventional Conjoint and found a validity lift of 25 to 85% in explanation power. Results show typically higher face validity and give seamless recommendation where pricing and product configuration naturally fit to positioning, segmentation and marketing strategy advice, as it is derived from one single data source and one single causal model.

p.s. More on NEUSREL and Universal Structural Modeling here.

Causal Analysis – the better Key Driver Analysis

By | Allgemein

Causal Analysis – the better Key Driver Analysis 

Marketers and market researchers who are hoping to find the most effective actions in data are confronted with dozens of statistical methods.

At some point someone started to rename regression-based techniques as “key driver analysis” in order make it more obvious what they do – figuring out what drives success. However this has not been ‘state-of-the art’ for decades. Companies can prevent expensive mistakes by switching from Key Driver Analysis towards Causal Analysis. Read Why!

The challenge with Key Driver Analysis

Key Driver Analysis has been a major step, helping to guide companies away from spurious correlations and dangerous mistakes. In fact, the term mainly describes regression analysis or similar econometric techniques and comes with its limitations.

Inflexible

Regression techniques estimate parameters of a model which represents the impact of certain factors towards an outcome. Within this model, it is assumed that factors are independent. This means that the parameter is always the same no matter what the situation.

E.g. a bottle of water leads to a certain growth of a tree, no matter how much water it already had and whether or not the tree gets enough sun.As the example shows, these assumptions can be very unrealistic. The techniques are not able to adopt assumptions to data and mostly don’t even consider the nonlinear nature of reality.

Only direct Effects

On top of this, regression techniques live in a simple world: factors influence outcomes – full stop. When a TV campaign drives Google Adword Views and those clicks then drive sales, a regression method would give the “TV factor” a small impact parameter because with Google Adwords  can perfectly explain sales.

In most real life cases, factors influence each other, just as TV drives Adwords. If you do not consider this, you are not measuring the full impact but only the direct impact.

Key Driver Analysis is simply outdated. UNILEVER first applied it in the 1920’s. It has been around for nearly a century! Companies need a methodologic upgrade to be able to compete in the 21st century.

Why a solution of Key Driver Analysis’ challenges would be so valuable

If we could only quantify the full, and not just the direct, impact, if we could model data with a self-learning system that finds realistic but unknown features such as saturation effect or moderating factors… we would be able to understand much more realistically what drives success or why customers choose a brand.

More importantly: We would no longer be forced to impose unrealistic assumptions or badly founded hypotheses. We would have more effective recommendations and sales and marketing actions. Increased advertising effectiveness will help save millions of marketing dollars.

The powerful solution: Advanced Causal Analysis

The roots of causal analysis are a hundred years old as well. It all started with experimental test designs. As well as the fact that many things in business are much too expensive to test, most real life experiments face another issue:

there are other drivers of outcome that cannot be perfectly controlled (e.g. with a perfectly randomized sample). That is why a kind of “Key Driver Analysis” is needed to quantify the impact of factors including the experimental action.

The classic causal analysis

In the 1960s and 70s, regression had been extended to path modeling techniques (Structure Equation Modeling SEM, Partial Least Squares PLS) which are modeling not just direct, but the full effect of causes.

Later in the 1990s, Baysian nets (or Direct Acyclic Graphs) had been developed which, on top of SEM and PLS, can (in simple networks) identify the causal direction between variables out of data.

The age of Machine Learning

All those methods have a common limitation: they are not self-learning systems, as researchers are forced to assume a fixed model – most of the time, they also need to assume linearity and independents of factors (remember the tree example!).

The Universal Structure Modeling approach implemented in the NEUSREL software changed that in 2008. As conventional causal analysis requires making assumptions that nobody can justify in a business context, USM is a true breakthrough for practical applications.

The good news: Advanced Causal Analysis is proven, reliable, tried and tested

Since 2008, there have been hundreds of business projects conducted with the help of Advanced Causal Analysis, Universal Structure Modeling and NEUSREL. Many reputed scientists have approved the methodology, dozens of papers have been published and highly reputed brands such as Audi, Deutsche Bank, L’Oréal, P&G, Unilever and many more are leveraging these techniques.

These companies understand up to 300 percent better why customers do what they do. They gain insight that was previously unknown. They explore nonlinearities and discover saturation effects and optimas in their data. They learn how factors and conditions interact and how target groups naturally divide into segments. They learn it without knowing or hypothesizing about it beforehand. This enables these companies to step on a learning curve never seen before.

Copy Testing: How Copy-Testing based on Causal Analysis Leverages Advertising Effectiveness

By | Allgemein

Dozens of studies have been shown that about 70% of advertising impacts cannot be reasoned by how much you spend nor which media mix you have chosen. Rather it is all about crafting an effective creative. But how do we craft an effective advertisement? It is only about 30% art but 70% hard science. This article highlights a proven Copy-Test-based method that ensures creatives that drive Advertising Effectiveness.

The Challenge with Copy Testing

Copy testing of advertisements is a common practice. Still the majority of campaigns don’t meet expectations. What’s wrong? Many interesting research methods have been added to the copy test toolbox in recent years. Eye tracking, skin resistance checks, Neuromarketing with EEG and fMRT measurement, facial emotion coding and Implicit Reaction Testing (IRT) are all powerful methods to detect to what target customer might think and feel about an ad.

But something is missing here

But something is missing: When trying to interpret one of those measurements you assume that you know how this measure relates to the outcomes you tries to influence. This is not the case. There is no single measure that will give success. Furthermore, how components interact with each other might be different per segment, product or situation.

Advertising can be stimulating but there are products where this characteristic is false. So far all surveys are built on the assumption how the measures are connected with the actual target.

Why a solution of the Copy Testing dilemma would be so valuable

If we could decode the hidden formula that enables us to predict outcomes like purchase intention or brand consideration using copy test data, we would be able to optimize the creative in a way that outcomes are maximized.

We would be able to write an individual manual of do’s and don’t’s for every brand on how to craft advertising. Such a manual would be our crash barrier on the creative road of advertising effectiveness.

The powerful solution: Copy Testing based on Causal Analysis

In a model-based copy testing approach, we need collect several information during a copy test:

  • Data about the direct impression of a copy test – newer measurement approaches like IRT and facial coding can play a vital role here.
  • Data about how the direct impression changes attitudes towards the brand and how it increases recall and recognition of the ad.
  • Data about the outcomes of interest, such as purchase intention, willingness to inform, brand consideration
  • Data about the respondent like demographics, their situation and history. Are they already customers? Are they a bargain seekers? Do they know prices well? All those items that might have impact on their desired outcomes.

The ultimate question is: how do we identify which changes would improve outcomes? It is a challenge to conventional methods as they are built to fit parameters of a known model.

What is needed is a method that builds a prior unknown model in the light a data, a model that incorporates indirect causal effects and do not neglect nonlinearity and interaction of those unknown relationships. This is exactly the strength of the Universal Structure Modeling approach that is performed by the NEUSRELTM Software.

The good news: Copy Testing based on Causal Analysis is proven, reliable, tried-and-tested

Results of model-based copy-testing has proven to be eye-opening of marketers and gives an unbiased rigorous guidance:  A recent study has looked at a TV spot of a beverage. The spot made a great impression; everyone liked it, only it did not do anything with consumers. We found that the spot was evoking nonproductive emotions and needed to switch from balance to stimulating emotions.

We found that the spot missed brand anchoring during the advertisement and especially at its end to leverage recall. The optimized spot broke all records and has become the best performing advertisement in the brands in history.

Some say, similar recommendation could have been produced by alternative approached. This big and fundamental difference to any other approach is: We know and have statistically proven that our recommendations are the main levers to maximizing market outcomes. Are other approaches able of the same? We replace “hope” with “knowing” with our selflearning causal analysis.

p.s. More on NEUSREL and Universal Structural Modeling here.

Case Study “Predictive Customer Management” – How Analytics Brings Customer Value, Needs-Segments and Campaign Effectiveness in One Go.

By | Allgemein

Customer Lifetime Value:

An automobile manufacturer encountered the Pareto phenomenon: A small number of the customers are responsible for a large part of the earnings. In practice, it is often not easy at all to be able to recognize what a customer possesses in terms of future value. On top of this question arises how to capture this value. To find an answer you need to think through the whole process of marketing and sales to its logical end. This is exactly what the automobile manufacturer did.

Step 1: In the rearview mirror, you don’t see the upcoming curve: a future-oriented calculation of customer value

Customer-value-oriented customer management is only in its rudimentary stages among automobile manufacturers. Sure, there are corporate customers and there are customers who buy the expensive types of cars.

But what will they buy in the future? How loyal are they to the brands and to the authorized dealers? How probable is it that their next car will be a new car? How much business will be a result of personal recommendations? How much business in service and accessories will the customer trigger? These are central questions to which, up until today, there aren’t any satisfactory answers…no: weren’t!

A research panel of customers delivered the basis for the data: customer profile data, personal car purchases, precise quarterly service expenditures. An additional survey among this panel delivered the information necessary to understand the sales triggered by personal recommendations and to also understand sales expectations over the long-term.

Predictive Analytics (here NEUSREL) delivers the mathematical formulas that can calculate, from customer profiles, the diverse, future revenues to be expected. The result: a forward-looking customer value for every customer, which is calculated based on a small set of variables from that customer.

Step 2: What to do with the valuable customers? Derive segment-oriented support concepts from the data.

“What to do?” The customer value tells you how much value can be “gambled away” when you don’t pay attention. It doesn’t tell you what you can do in order to retain the value or increase it. Here is where Predictive Analytics can help too. Because approaches like NEUSREL also deliver the importance of the variables. And these tells you what the driving force of substantial customer value is.

The automobile manufacturer found out, for example, that aficionados of high performance cars not only invest significantly more today and in the future in production, but also most notably in service. A substantial maintenance division that takes care of high performance cars was born.

Furthermore, the company found out that brand loyalty and trust explained almost half of the customer’s value. The segment of the emotionally uncommitted provides a segment that can be excited by measures tailored to individual customers. Their affiliation can be tested with just two “killer questions.”

 

Step 3: Does the customer respond to my actions? More effective guidance through prognostic Action Scores.

It is a decisive step when realizing which buttons you can push on a target customer.  Nevertheless, it can happen that the chosen measures don’t take effect. Perhaps because the person, per se, doesn’t read any emails or advertising letters or maybe isn’t receptive to any discount offers.

What you need is an assessment of whether the call, the letter or, for example, the discount offer will actually evoke a positive reaction. In this way, you can avoid having your marketing and sales investments end up in your customer’s “spam folder” and their effect going up in smoke. You don’t want to run the risk of putting your customers off.

How does that work? The targeting, with the help of Predictive Analytics, is a practice that has already been successfully used in direct mailing for over 20 years. On the basis of customer characteristics and customer reactions from the past, an Action Score can be calculated, which provides a probability of a customer’s reaction to a certain measure. With the help of Action Scores, the automobile manufacturer increased the effect of the measures taken by 84% – with the same investment.

Step 4: Change is the only constant: the necessity of a control process

What about when the planned measures are new and therefore there aren’t any empirical values available? Then you need a pilot phase for these measures. For this, customers will be addressed randomly. After a few weeks, a Predictive Model can be built and the measures can be efficiently guided based on the Action Scores.

The value of a customer will also change over time, just like their needs and the products of your company. That’s why the automobile manufacturer decided to conceive the issue as a control process. The customer value formulas are regularly calculated anew. Different than the conventional segmentation approach, the needs segments are also regularly reexamined and, as necessary, adjusted and brought to life with Action Scores.

p.s. More on NEUSREL and Universal Structural Modeling here.

Getting More from the Brand Tracker: The SONOS Case.

By | Allgemein

Most companies have one. A regular measuring instrument to assess the brand’s status quo. But more questions than answers are often thrown up by the large number of facts measured. How can we enhance brands and increase sales? How should positioning be sharpened up, how should be make communication between touchpoints or segments different from each other? The SONOS Case shows us.

SONOS is a home audio brand that has established the new multi-room audio category with its active Wi-Fi speakers connected to the internet. Until very recently, there was only one maxim in SONOS’s marketing: Awareness at any price – retailers will do the rest. The advertising was as expected– shrill and loud.

Here is an example:

Global market research set itself the task of getting to the bottom of the brand drivers. Conventional driver analyses threw up questions, did little to explain the variance in customer behaviour and were not very plausible in their restrictive assumption. SONOS therefore approached Success Drivers. We used the available data to build up a Universal Structural Model (USM). The results were eye-opening.

It All Depends on the Method

Comparison with a correlation analysis showed, for example, that a positive evaluation of Customer Services has a strong correlation with the intention to purchase – however, USM shows that service is not the cause. Consumer who already are already customers tend to assess not only the service more positively but also have higher likelihood to purchase even more.

Comparison with a conventional driver analysis (regression) shows that an association of the brand as a “Leadership Personality” directly and negatively reduces the intention to purchase, but the same driver increases identification with the brand. Since the latter has a great influence on the intention to purchase, “Leadership Personality” is therefore not negative in the overall effect. This is a finding that only path analytical approaches would reveal.

Comparison with conventional path analyses shows that a universal modelling approach models non-linear and interactive connections and, in the case of SONOS, is 60% better at explaining why customers buy. In particular, it because obvious that a perceived “Relax” emotion was barely significant in the linear model, but was extremely important in the non-linear model.

If the brand is seen as hardly relaxed instead of moderately relaxed, consideration and intention to purchase fall dramatically. Moreover, in certain subgroups a high level of relaxed feeling was a genuine sales booster – in methodical terms we call this phenomenon an interaction: drivers can have different significances and importance in various contexts or segments.

Key Insights are Taking Effect

In conclusion, SONOS gained so many insights that completely turned the marketing around. We noted that Paid and Earned Media need completely different communication as opposed to Owned Media or at the POS.

We discovered that SONOS was somewhat neglecting the basic lever of “Sound” as a category in its communication. A new Premium Speaker has been developed and target customers are now seeing much better that the unique SONOS features (multi-room, software updates, etc.) are the reason for a unique sound experience. Every creation has been optimized so that there is always a relaxing feel. See for yourself:

Why are our TV ads not as effective as they should be?

By | Allgemein

How does the classic copy testing go nowadays? Recall and recognition are recorded and questions are asked about how the film was assessed and the message understood. Is this enough for your marketing department to optimize the copy? No? You are not alone. The test is sufficient for an evaluation, but all too often there are only a few indicators of how the advertisement can become more powerful.

Two things are missing

Firstly, we have to work out how to measure a measure of success! Recall is not success per se, as the long-winded example of Benetton shows. Intent to buy or brand preference are better indicators of success. Secondly, we need to take an analytical look at the drivers of success.

Analytical look? How is that supposed to work? You can have the contents of the film dissected by communication experts or have volunteers undergo a brain scan. If you are put off by the costs of this, there are tried and tested ways of collecting suitable data for a downstream cause analysis by means of online surveys.

This is how it’s done!

You show a sequence of advertisements and randomly switch your spot in it. After a distraction, you measure recall and recognition, you measure the relevant set and the intent to buy for your product and have the product assessed in more detail. As well as items that measure the message, unconscious emotional aspects should be quantified (e.g. with reaction time measurements). In other words, you are measuring what is changing into potentially relevant drivers of success as a result of the spot.

But the art lies in reading the data! Because if an ad is good, the perception of “stimulance”, recall or brand trust may improve, but which of the three have an effect on success?

Don’t the various components influence each other? Does a “stimulance” emotion, maybe enhance recall, and recall increase brand trust? Maybe, but only to a certain degree – maybe “the more, the better” attitude from linear analysis models distorts the true effects?

Don’t worry, of course there is a tried-and-tested method that takes account of all these uncertainties. And I bet you already suspect which ones…

p.s. More on NEUSREL and Universal Structural Modeling here.

Survey and Primary Data are Getting Married – A Dream Wedding for Key Driver Analysis.

By | Allgemein

Survey and primary data are getting married. This is a report from three weddings that made an impression.

Profit Driver at Retailbanking

A bank wanted to find out what the drivers were for additional business with its customers. They linked survey data on customer satisfaction and further recommendations with customers transaction data, contract conclusions and gross premiums. This was possible underdata protection legislation because anonymized data sets were linked and further processed by an external agency. The result of the cause-effect key drivers analysis was unexpected, but more than plausible:

although customer satisfaction is an important basic condition, its contribution to contract conclusions evaporates for the simple reason that policies are taken out only rarely – every 5 to 10 years on average. It became apparent that customer satisfaction is not a suitable panacea – since an increase in the use of the scattergun approach offers only a low ROI. The result is a marketing and sales concept focused at customer touchpoints at and just before the time a customers need arises.

After-Sales in the Automotiv Area

A automobile dealer was interested in the drivers of its profitability. Using operational indicators, data on sales campaigns and its survey data, hundreds of dealerships were characterized. The findings from the cause-effect key drivers analysis were manifold.

Properly carrying out maintenance work turned out to be an important driver of satisfaction, further recommendations, revenue and cost savings. Instead of an abstract importance of around “0.5”, a concrete effect on profits could be assigned to it: if a typical workshop achieves the performance for “Proper work” of the best workshop, its profit margin rises by 15 per cent for this reason alone.

Distribution of a Consumer Electronics brand

A kitchen appliance brand wanted to identify the effect on profits of sales campaigns and, for this purpose, formed control groups of shops that did not run any campaigns for several months. An initial model shows no impact for most campaign instruments like trainings or calls. No wonder, because after all modeling just uses turnover data to measure short-term effects on turnover. It neglects long-term effects that result from the campaign actions that gradually improving sales performance in the long term.

The brand supplier integrated the key data from the mystery shopping into the description profile of every shop every month and as a result massive indirect effects soon emerged in the cause-effect key drivers analysis. Certain campaigns not only increased the mystery shopping category of “Activism”. This category also proved to be extremely good at boosting sales. Suddenly, “No effect” because a demonstrable “Baseline effect” that the company could measure in monetary terms.

Three examples, one finding:

your survey data are valuable. But your customer, turnover and transaction data give them the required polish.

– Frank

p.s. More on NEUSREL and Universal Structural Modeling here.