If your company is using analytics to improve any aspects of its sales or operations, success can mean the difference between using or not using a specific variable or data set within a model. Unfortunately, using all the data within an enterprise is nearly impossible. Many companies spend millions of dollars trying to enable such a strategy, often abandoning the process due to cost and time required.

Given these challenges, it is crucial to prioritize which data sets are included in model building and testing. Basic quantitative variables, like demographic information or financial numbers (i.e., the “usual suspects”) always make it into testing, but some types of data are often ignored. Below is a list of five types of data that are ignored by many companies when running customer analytics projects, yet are likely to yield valuable insights and drive model performance.

#1: Transactional Data

Companies that sell products or provide services track the transactions they make, yet they often limit customer analysis to financial variables. Go beyond metrics like “average revenue” or “customer lifetime value” and use transactions to understand what sort of customers you actually have. Transactional data helps segment your customer base into groups with specific tastes, preferences, and loyalties.

#2: Customer Correspondence

Most customers leave feedback in the form of e-mails or voice messages. Dealing with unstructured, non-numeric data like this comes with a steep learning curve but is the only way to analyze how happy or upset your customers are. A common alternative to such data is surveying your customer base, which is expensive, time consuming, and needs to be repeated on a regular basis.

#3: Other Model Results

Has your company built models before? Model results are typically kept apart from the “core” customer data, yet can provide important insights into customer behavior; leverage your analysts’ prior experience by including older results in current modeling efforts. A case in point is credit scoring: these are the results of advanced modeling techniques, and often provide significant improvements when added to other models.

#4: Product Descriptions

Like customer correspondence, descriptions and photos of products are contained in formats that are difficult to analyze. At the same time, they are the core sales strategy for any consumer-facing company: it is the product descriptions and photos that excite and entice customers to make a purchase. Analyzing what works, what doesn’t, and how it can be improved can significantly improve sales and conversion rates.

#5: Employee Engagement

If you’re a consumer-facing company, your employees are likely buying from you (and if they’re not, ask yourself why!). Indeed, they should be your biggest advocates. Survey them, analyze their product usage, and involve them in product development and marketing. Ask if you can track their movements on the consumer-facing side of your website, and interview them to understand their purchasing habits. Use this to improve how you build your models, segmentations, and ultimately, your sales campaigns.

Have you worked with data types outlined above? Share your experiences below or contact us to learn more.

Canopy Labs founder and CEO, Wojciech Gryc, was interviewed on Canada’s Business News Network last week. The interview focused around how the Canopy Labs platform can be used to help small and medium-sized businesses analyze their customer data.

NetFlix recently made headlines by announcing it would not be using the winning algorithms from its NetFlix Prize competition in production. The NetFlix Prize awarded $1 million to the team that improved movie recommendations by 10 percent over NetFlix’s proprietary algorithm. Paying $1 million for an algorithm that you don’t use might seem counterintuitive, but in reality, this illustrates the complexities of using analytics in an enterprise setting.

Enterprise analytics is not academia, though this truth is often hard to bear, particularly for (younger) scientists entering the corporate world. The Canopy Labs team has seen numerous enterprise analytics teams and projects, and here we share three observations that drive our thinking in analytics within the enterprise, and how it differs from more academic settings.

#1: Model performance is only important if it drives real-world results

Model comparisons are what drove the NetFlix prize and similar competitions. Internal analytics teams often compete against a benchmark, be it internal or external. Analysts sometimes argue over model accuracies or other success metrics, and show how one model’s performance might be more significant than another.

Statistical significance, model accuracy, and other metrics do not matter as much as real-world results. True, a more accurate model might be more likely to achieve better results, but the relationship between higher accuracy and higher sales or conversion rates is rarely clear. A key challenge for data scientists is to find better ways to estimate how theoretical model improvements will actually benefit real-world applications.

#2: Research lives within a budget cycle

Model development and other enterprise analytics live within a budget cycle that needs to be approved, and budgets are often limited. Building highly effective models often depends on:

1. External data sets to complement whatever internal data your company has (not cheap).
2. Advanced software to run new types of models (moderately expensive, particularly if you include training costs).
3. Specialized hardware to run complex and computationally intensive algorithms (very expensive).

Often times, model effectiveness is balanced by the business need. A credit card company optimizing its fraud detection models can make a clear argument as to why it should invest in supercomputers and new modeling tools. Unfortunately, this might be less true for smaller retailers striving to grow sales by 2%.

#3: Many people go home at 5pm (and other time constraints)

A limitation worse than that of the quarterly or annual budget is that of people’s time. Models are not built in isolation: other priorities get in the way, particularly if staff need to decide between experimental ideas and priorities for senior executives. Very few companies can actually afford full-time modeling staff and even then, incentives are rarely structured to encourage working on models beyond a set time frame or level of accuracy (which only perpetuates the challenges outlined in point #1).

So what to do? We don’t write about challenges without thinking about solutions. Building better tools is one option, but more importantly, it’s important to automate the process itself. Today, too many resources are being expended on baseline models rather than truly innovative approaches. Give people more time and energy (by making it easier to reach those baselines) and you will get more innovative models.

Making this possible is a software development problem — one Canopy Labs is working on. We’ll describe our solution is an upcoming post.

Analytics is not user friendly. Most statistics packages, machine learning tools, and related platforms were created by scientists, for scientists. There are huge benefits to this as a result: the platforms are programmable, versatile, and often very quick when it comes to solving difficult analytic problems.

Over the last few years, however, the analytics industry has changed. Today, many non-statisticians and non-scientists base their decision-making on mathematical models and statistical analysis. Like most people in the industry, I welcome this change and am excited to see what innovations will come about as a result. The challenge, however, is ensuring these analytic tools are usable by people without complex training or statistical backgrounds.

Ask yourself: when was the last time someone discussed the wonderful user interface of their preferred analytics tool? In my case, never. Instead, I hear numerous complaints, and they focus around three core challenges around user experience and analytics platforms:

1. Outputs with too many (or irrelevant) details. Very few people actually understand or care about analytic aides like P values, standard errors, R² statistics, etc.
2. Inability to use results for specific actions or activities. For example, some tools will give you high-level results around your customers (e.g., a model) but won’t give specific predictions tied to individual customer IDs.
3. Complex and confusing scripting languages required for any complex work.

These challenges are not a result of poor design of existing systems, but rather the unique needs of a new user base. This leads us to a unique approach to developing our own platform, through three key design requirements for new platforms and analytic tools:

1. Cater to the specific needs and activities of individual user groups. Rather than building all-purpose statistical tools, there is an opportunity to focus on specific needs and functions of end users. For example: accountants, actuaries, marketers, and advertisers all use logistic regressions, but do so in different ways. Cater the experience to each unique use case.
2. Understand the level of analytic rigour needed by users. For example: academic researchers worry about P values, standard errors, R² metrics, and so on. Many people do not. Reframe model results to speak to the analytic rigour/experience of your end users.
3. Develop modelling tools that are customized to the specific needs of users. While existing quality metrics (e.g., R², root mean squared error, etc.) are useful for comparing models, they become confusing when discussing applicability to real-world scenarios. If my model’s R² metric is 0.73, how can I expect it to perform in forecasting sales in my stores next week? Reframe or rebuild analytic models to meet the needs of the (non-academic, extremely applied) end user.

These are formidable challenges and much needs to be done in the coming years. It is exciting to see what future analytic tools will look like, but one thing is certain: the analytics landscape will be significantly different in 2 or 3 years.