Prior to Time Inc, Tamara held leadership positions in consumer analytics for Stuart Weitzman and Gilt Groupe. She is passionate about the application of analytics across a wide range of business problems and is excited about the opportunities presented by big data. She is currently the Vice President of Strategic Services at Action IQ, an enterprise customer data platform. There she works with companies to help them make data-driven marketing a reality. Below she teaches us how to maximize our data to its fullest potential.
What is the difference between predictive modeling and machine learning?
Machine learning as a concept allows you to scale the development of predictive models and form systems that can learn from data with minimal human intervention. That takes the pressure off one statistician building the model, implementing it, gauging the results, looking at the model's performance and adjusting the model based on the results. Instead, it allows machine learning systems to carry out that process automatically.
What is hypothesis testing, and how can it be used in a business context?
Hypothesis testing is a very useful tool once you understand it. It allows you to put a lot of ideas into the market and capitalize on opportunities with great confidence. Hypothesis testing starts with understanding your business's default action in the business context. The default action is what you're currently doing, what you've been doing, and what you're comfortable with it. Hypothesis testing continues by testing an alternative action that you believe could be better for your business.
For example, say you send your daily emails to your customer base at 10 am every day. That is your default action. You already know your average open rates for that send time. An alternative action could be sending your emails out at noon instead. Once you have sent emails at that send time for enough time, you will have the necessary data to compare open rates.
Hypothesis testing calculates statistics based on collected data and then decides if you want to stick to your default action or switch to the alternative. That's where statistical methods come into play. It's not enough to collect the data; you need to interpret it.
Hypothesis testing calculates statistics based on collected data and then decides if you want to stick to your default action or switch to the alternative.
How are the results of statistical methods interpreted?
When interpreting the results of statistical methods, essential terms to know are type one or type two errors. Type one error, sometimes called significance level, is when you stay with the default action. A type two error occurs when you go with the alternative action. By design, type one errors should be more painful than type two because switching your action always assumes some risk. If you decide to switch from your business's usual, you need to be very confident that your alternative action will produce better results.
What kind of support is needed to present data to business stakeholders in order for them to understand it, process it and make informed decisions?
Business intelligence support would be extremely beneficial when data needs to be presented to business stakeholders. In addition to Excel and SQL tools, there are a variety of business intelligence tools that help with this storytelling. The right tools can help analysts evangelize this information and get buy-in from business stakeholders.
Once stakeholders have begun consuming the information, brainstorming new ideas and seeing how the information can become actionable, the business can move into the testing stage. The testing stage is where you need a statistician or somebody familiar with the statistical methods of hypothesis testing that would allow you to test out different hypotheses and give you a comfort level with the results.
How can data be used to optimize your business?
As you further your levels of sophistication and begin using predictive modeling, you can optimize your business. A granular level example would be optimizing your customer's experiences through email marketing.
Instead of testing something for specific groups of people, you can attempt to predict the best time to send an email to every customer based on their profiling behavior. Your gains from this model are much more significant because you're optimizing based on every individual customer on your file.
When you begin to scale those sophisticated solutions, you are efficiently utilizing machine learning without employing a lot of people to do the manual work. Data analysis, statistics, predictive modeling and machine learning, can be super impactful across different functions and industries.
What are the most important elements needed to draw the right conclusions from data?
Two necessities are accurate data analysis and somebody who understands statistical methods within the context of your business.
Data analysis does not happen in a vacuum. Statistical and testing methods need to be applied to very specific business problems. So, it's important to understand your context, how data is structured and what each field means in your database. You need to be able to address questions like: How do you define a valid order? Or how do you define an active customer? It is very plausible for two people from different departments to look at the same data and draw different conclusions.
For accurate reporting, the priority must be consistency around data interpretation. The next priority is to ensure the data collected is specific to business goals and uniformly inputted into the company's database.
Two necessities needed to draw the right conclusions from data are accurate data analysis and somebody who understands statistical methods within the context of your business.
Is it better to focus on predictive analytics before moving to machine learning (ML)?
ML is enabling your predictive analytics at scale. So, yes, it would be beneficial to test the impact of predictive analytics for your particular problem before moving into building machine learning structure frameworks.
What is a statistically significant sample size? And what is a minimum time frame necessary for leveraging a data set for predictive analytics?
Two thousand respondents is a robust sample. There is an opportunity to analyze customer behaviors within that sample size.
I would caution to only do things future based when the environment is consistent. Prior to the pandemic, it was smart to look at the prior 12 months to predict the upcoming year. Now it's wiser to be a little more cautious and creative when future planning.
How do you upskill junior talent into pulling insights from analytics?
Mentorship is super important. I think it's challenging when you only have somebody junior on your team who does not have a connection with someone senior. If there is nobody like that within your organization, I would encourage participation in outside groups. There are a lot of organizations, meetups and communities that get together. Professional development classes are also beneficial, but hearing from people with real-life experience is invaluable.
Can you suggest some online courses on data analytics?
There are great courses on Coursera and Udemy.
