{PODCAST} In-Ear Insights: Predictive Analytics at BACon 2019

In this episode of In-Ear Insights, listen to the full, un-abridged audio from CEO Katie Robbert’s talk at the Business Analytics Conference (BACon) as she walks through use-cases for predictive analytics in multiple industries, from forecasting real estate sales to customer service call center staffing. You’ll also learn what constitutes good data and the overall predictive analytics process.

[podcastsponsor]

Listen to the audio here:

Download the MP3 audio here.

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

So today we’re talking about time series forecasting, which is really predictive analytics.

And so the general disclaimer slide, as mentioned in the opening remarks this morning, you probably seen this a bunch today.

So introductions, I came rubber, I am the CEO of Preston sites.

I’m also a proud partner and part of the pm square data science team.

I’m a noted keynote speaker and an IBM Business Partner.

So as we’re talking about predictive analytics today, one of the things I want you to keep in mind is that it is not going to tell you exactly what to do, it is meant to push you in the right direction.

So this is a really great quote, that says the goal is not to predict the goal is to change behavior to change outcomes.

And so keep that in mind, whether you’re, you know, going through this presentation, or doing predictive analytics on your own, is meant to sort of guide you into the right direction to make a more data driven decision than just guessing which we’ll get into.

So before we get into what is predictive analytics and how to use it, I want to just quickly go through the data analytics hierarchy.

So you read this from bottom to top, the bottom being the foundation of your descriptive data, sure quantitative data, what happened, the problem is that most companies are stuck at this bottom layer.

And they can’t even really tell you what happened.

They have a hard time collecting consistent data, data that tells them what’s going on.

And then even making sense out of it reporting, we call that dark data.

So a lot of companies are sitting on these piles of piles of data that they can’t make sense of, so they can’t even really figure out what happened.

Next is your diagnostic data.

So once you know what happened, you need to understand why it happened.

So these might be your customer feedback surveys, your employee feedback surveys, any sort of market research that helps you paint the picture as to why people made the decisions that they made.

And then next, once you have those two auto wrongs, you can then safely move on to predictive analytics, which is the what will happen and that’s always that most today.

Once you have those, you move on to the prescriptive, or what am I going to do about the thing that’s likely to happen? So those are your actions.

And then I did in an ideal world, where a few years away from this type of deep, deep learning, you can find all of those together, your descriptive, your diagnostic, your predictive and prescriptive, and a deep learning model will take all of those things, and then spelled the opposite say, here’s what’s going to happen.

Here’s what you which will be fantastic.

I look forward to that day will save me a lot of time.

So why should you use predictive analytics? Well, it seems pretty straightforward, but we’re going to go through it pretty quickly.

So first and foremost, guessing is bad.

You might be saying to yourself, Well, I don’t guess my make decisions.

But one thing that I want to know that I’m guessing is that if you’re not using your own data to make the decision, you are likely introducing some sort of an unconscious bias.

So think about a real straightforward example.

Let’s say you own a pizza shop, and you doing inventory and you know, in your mind and your heart, your stomach, you love to eat pizza on Sundays.

Well, the data actually tells you that people tend to order more pieces on Wednesdays, but you’ve not front loaded for Sundays, because you like because in your mind, that’s when you want the pizza.

So it’s just something to keep in mind, you may not realize that you’re guessing, but you may be introducing some sort of an unconscious bias.

The next reason is, again, you’re only looking backwards if you’re using your descriptive data of what happened.

So you can’t change what happened in the past, there’s not a whole lot you can do about it, it can do some informing of what’s going to happen in the future.

But it’s not really going to tell you an indicator because it doesn’t factor in other variables such as trends, what might come up new information.

A predictive forecasting also tell you where to spend your money.

So when to step up, when to step down, when to spend money on ads, when not to spend money on ads, the worst possible thing you can do is just start throwing money at everything, and hoping that sticks, because that rarely ever works out.

And then the last reason why not the only reasons but the last one in this presentation is to get ahead of your competitors.

So it gives you that competitive edge.

If you’re using your predictive forecast to make decisions about your business, then you can get ahead of your competitors who aren’t doing that, because it’s not something that’s commonly used by a lot of companies.

So what the heck is predictive analytics? Well, there’s been a lot of talk about it today.

It’s a type of machine learning.

There are four different kinds of machine learning and predictive analytics, all within the continuous supervised meaning you have your own metrics, and you have a known outcome.

The other three are continuous, unsupervised, categorical, supervised and categorical unsupervised.

For the purposes of today, we’re not focusing on those, we’re only focusing on that one bucket of continuous supervised where you have your two nodes.

I hate this formula, because it scares me.

You don’t have to know how to calculate it.

But this is the formula, this arena model that goes into predicting a time series forecast, you just need to understand how it all works, you don’t actually need to be able to calculate it, the machines will all do that for you.

So I’m going to walk through a very straightforward example of how to think about an agreement model with seasonality.

And that’s the US.

So think about the last time you drove somewhere, you put your destination into Google Maps.

And Google Maps said, on average, that’s the average part, it’s going to take you using this route about 15 minutes.

And it’s using historical data based on everyone else who’s ever taken that trip, and it’s collected all of that information.

Now, this is a map of Boston, and if you’ve ever driven in Boston, which is where I’m from, it’s a nightmare.

And so it’s telling you on average, 15 minutes, but what that really means is factor in like another 45 minutes, and we’ll get into the wise for that.

So that’s your average in the remodel, then you then you add in the moving part, and that’s your speed limit.

So on certain parts of the road, the average speed might be 65 miles an hour, it might be 60, you might be 30, you might get into a slow down, which causes you to only go 10 miles an hour.

So all of that data then has to factor in to your final output of how long it’s going to take you to get from point A to point B.

And then since you’re in the city, you have an accident, you have construction, you have all of those different roadblocks, you have so many drop it back and street thing.

And so that’s your auto regressive fees.

And so you then have to factor in data that you weren’t currently aware of into that model, which is now going to adjust the amount of time it’s going to take you to get from point A to point B.

So that 15 minute ride that seems really super easy is now 45 minutes an hour.

Well, guess what? The accident square guy picked up his back, he figured it out.

So now you have to readjust your model for even more new information.

So the integrated pieces, it’s constantly taking a new information to figure out how justice How do I give you a more accurate output for how long it’s going to take you to get from point A to point B? Well, guess what? It’s Boston.

And even in Chicago, you have terrible weather, especially in the winter, it might snow, it might drop five feet, and you can’t go anywhere, you might have that sun glare the certain time of year you can’t see anything.

And that’s the seasonality.

In this instance, when you’re thinking about seasonality in your time series model, this might be time of year, it might be end of quarter holidays, when people’s budgets were low for those types of things.

That’s the seasonality that this sob model will take into consideration when it’s projecting forward what might happen.

So that’s the underlying structure of the time series model.

You can actually run the time series model now predict forecast in IBM.

It’s the IBM IBM SPSS modeler for predictive and it’s found right there.

If you haven’t tried it, I recommend playing around with it.

There are other ways to run a predictive 14, depending on how big your project is.

I like to use we built our own custom code in our which is an open source developer software.

But it essentially does the same thing because it’s all based on the SPV model.

So I want to walk through some examples of how to use a time series forecast in your day to day.

And I think that this is really important, because a lot of times people get this notion that predictive, it’s like it’s too far out of reach, I can’t do it.

It’s not something I can do myself.

But these examples, I’m hoping are fairly straightforward.

And like, okay, I can see myself doing this, or starting to use this.

So we took a look at the real estate market, customer service calls, job search and content planning.

So before I get into that, is anyone here familiar with spurious spurious correlations? Okay, couple of people.

So if you’re not familiar with spurious correlations, essentially, what it is, is this website, and I believe you’ve now been turned into a book where this author has taken two different data sets and shown how they match up together.

Well, just because two datasets match up together does not mean that they are related.

It’s not, it’s not indicative of causation.

So for example, and the number of people who drowned by falling into a pool does correlate with the films that Nicolas Cage appeared in.

But that does not mean that one cause the other.

So I do want to just sort of put that little disclaimer in, as Just be careful when you’re running a time series forecast, because it may not truly be indicative.

If your correlations match, you really have to sort of step back, this is where that human part of the whole process really factors in is using some self judgment.

So when we looked at the real estate market, before we ran the time series forecast, we actually did a correlation between search intent of people looking for homes near me homes for sale, using Google Trends, data, and correlating that with Zillow data, the data is the data set was from 2014.

President.

And what we found was that the correlation between the two datasets was strong enough that we can with confidence, say that people looking on Google for homes for sale near be is indicative of the volume of homes that will actually be sold.

It’s not just a random coincidence.

And so this is for the state of Illinois.

And it’s interesting, because what our time series forecasts saying is that perhaps we’re at the top of that real estate bubble, and now the volume is going to be going down for the next couple of years.

So if you’re a real estate, this might be something for you to start to look into.

Why is my volume going down? You know, where’s the new developments? Why are people moving my army? And those sorts of things, that’s what you can do with analysis like this.

The next is customer service calls.

So if you’re running a call center, if you have a product or service, and you have customers calling in, you probably want to know, well, why don’t they like to call most What do they like to call the least.

And so this is something that you be able to correlate with your product releases, or new announcements, those types of things.

But it’s helpful to sort of get a sense the head of when things are going to spike, and when things are going to drop off.

And this is helpful for staffing for training for professional development for onboarding.

And one of the things that you’d be able to do with an analysis like this is that starts to pull out things like sentiment, what are they saying? Are they happy? Are they sad, upset, are they leaving when the volume spike, so when it drops down? So that’s a really good use of a time series forecast in that occasion, because you can then figure out what’s good, what is likely to happen with our customer satisfaction, are they going to stay with us.

So the next example is job search and unemployment.

So again, a correlation between search intent using Google Trends, data of people looking for jobs near me open jobs, those types of searches.

And we ran a correlation with the Bureau of Labor Statistics data set.

And what we found was that there was a strong correlation between the quits and hires, excluding layoffs, there wasn’t a strong correlation there.

And it was strong enough for us that we can say with confidence that people looking for jobs is indicative of whether or not you’ll be able to find qualified candidates.

And so what’s interesting about this, and I didn’t put it in note is that people are likely to be searching for jobs more and more into mid next year.

And so that is potentially don’t quote me on this.

But it’s potentially indicative of a possible recession moving into next year.

So this would take a lot more research for us to really confidently say that, but those are one of those things that hiring managers, HR professionals who probably want to keep an eye on.

And this is the on aggregated, unemployment job search, you can see the term that we actually use, so indeed jobs, jobs, hiring your me jobs, your me work from home jobs, and all of the trends are roughly the same, saying it’s just a different level of volume for each of those things.

And then the last example, is content planning.

So with content planning, maybe you’re responsible for creating thought leadership content for original content, this was actually a calendar that my team uses.

And it’s stacked with keywords that we want our customers to be able to find a sport known for.

And what we found is that, starting in September of this year, a lot of people are going to be looking for a term such as Google Data Studio, and what is data science.

And so what we do with this type of information is we then audit all of our content to make sure that we answer basic questions around those terms.

How do I use Google Data Studio? Where do I find it? Where can I get it? When do I start, you know, all the general helpful questions, we make sure we have all of that.

And then what we do is take it a step further, make sure that content is optimized, it’s up to date that’s formatted correctly.

And then based on the trends of the weeks, we can start sharing hearing it organically, we can activate it, we can put ads behind it to make sure people are truly finding us for those terms.

And so this helps us figure out how to resource what to write what money to spend.

So where can you go wrong with the time seems forecast.

And so you can go wrong with a lack of planning, with a lack of data quality, a lack of follow through and having trying to do too much too soon.

So with the planning piece of it, this is probably for most people, the most boring part of doing any sort of project is the actual figuring out what do we want to do? How do we want to do it? But if you’re asking me, it’s probably the most important because it saves you so much time in the long run.

So first and foremost, you know, I encourage you to experiment.

But if you’re running a true predictive forecast that you’re going to use a measure, you need to have a plan.

So you want to set the strategy, have a goal, have your desired outcome, what is the question that I am trying to answer with this predictive forecast, then you extract the data.

Now we’ve used a blend of proprietary data and third party data is, this is where people often get caught up, they feel like I don’t have the right data to use, I don’t have data that I can run a forecast on, I would encourage you to start to look at third party data sets, like the Bureau of Labor Statistics like Google Trends, like statistics, like data doc up, there’s a lot of really good clean data sets that you can start to experiment with to see, does this answer the question for my company, will it help us move the needle forward to what we’re trying to achieve? Then you prepare the data so you clean refine and prepare the data set.

What you may do in this section is actually find that it’s necessary to do something that’s called feature engineering.

feature engineering is when you’re creating additional variables to answer questions.

So let’s say for example, your data is clean, it’s broken down by weeks, but what you really want to understand is the day or the hour.

So you you need to feature engineer additional variables to start to get to that level of granularity within your data set.

And this is the stuff that you do that, then you identify which variable to predict.

So you wouldn’t necessarily run a time series forecast against Facebook engagement if you’re trying to figure out revenue models.

So you need to make sure that again, it’s aligning with your strategy, and what’s the customer you’re trying to answer.

And then pretty straightforward, you create the prediction, and then you build a plan of action from the forecast.

So where else can you go wrong data quality.

So this is what we call the six C’s of data quality, you have to have clean data.

So prepared well free of errors, you need to get any sort of extraneous characters or anomalies out of your data set, it needs to be complete, no missing information.

So if you have five years of data that you have one week that happens to be missing, it’s not a complete data set, and it will throw off your full time series forecast, and used to be comprehensive, so it must cover the questions being asked.

So again, think back to that.

What’s the question that I’m trying to answer? Again, you wouldn’t necessarily pull Facebook data if you’re trying to answer revenue questions or financial questions or staffing questions.

So not any data set will answer the question.

It needs to be chosen with a purpose and then chosen data.

So nowhere relevant or confusing data.

So you may pull a data set that has a lot of extraneous information in it, that really has no bearing on the question that you’re trying to answer.

It also needs to be credible.

So one of the things that people fall short on and this is a whole separate conversation is analytics governance.

And so the governance piece is who owns it? who cleans it? How does it get updated, those types of things? How is it collected, and a lot of companies don’t have that level with their own data, so they don’t have the confidence in it.

If you’re using third party data sets, any good quality third party data set should have documentation around the credibility of how they’re collecting data around their analytics process.

And then lastly, without it being calculable, these first five steps become irrelevant.

So it needs to be calculable, must be workable and usable by the business.

So it needs to be in the right format, it needs to be something that you can directly import into your machine learning models into your SPSS model, whatever the methodology is, and then a lack of follow through.

So let’s say you put all of that time and effort into building your time series forecast, you go plan your strategy, you create the action plan, and then it just collects dust on the shelf.

That happens a lot happens with all of us happens to me, you know, you get really excited about this thing.

And then nothing happens.

So whether that’s resources time, whatever the thing is, and so that unfortunately happens.

And then you can’t measure it, you don’t know what happened.

And then you still sort of go back into those old habits of guessing of, well, we’ve always kind of done it this way, let’s continue to do it this way.

And then too much too soon.

So there’s this phrase that I kind of look kind of hate called this, boil the ocean, you can’t boil the ocean, you can only do one little cup at a time.

And I would recommend starting that way where the time soon as forecast, do a proof of concept, do something small so that you can see what it looks like to use that type of data to inform your plans.

And then you can demonstrate to your managers here C suite to your clients and customers.

This is what it looks like when we use this type of data.

These are the results.

So I would recommend not too much too soon.

But start with a smaller project, we can really control every piece of it.

So what can’t you predict? I want to be clear about this because a lot of the items on this list, there are ways to do some predictions on these things.

But using this particular methodology, this simple time series forecast, you cannot predict these things.

So things that are unpredictable things that have never happened.

Trying to forecast a presidential election with a time series forecast is a bad idea.

There have been no two identical presidential elections in order to give you a rich enough data set to predict the outcome.

There are other more sophisticated models, but this one is not it.

Things that have too many inputs.

Stock Market.

For example, if people had figured out how to predict the stock market using this time series forecast, I would not be here I would be on a private island, I would be elsewhere I would forget all of you exist.

If you could predict using that time series forecast, people would have figured it out already people are trying to figure it out.

And then lastly, you can’t predict anomalies.

So on September 10th, nobody could have said what was going to exactly happen on the 11th.

There was no time series forecasting model that could have told you, we’ve gotten better with some types of predictions, early indicators.

But this particular time series forecasts cannot do that.

Same thing with housing market crashes, stock market crashes, things that affect the market sort of f4 sets.

So I give you just a couple of customer stories.

So really sort of like pull it together.

So you have a better understanding of how my team has used time series forecasting and some of these third party data sets.

So we were commissioned by a very large social media company, you’ve likely heard of them, if you’re not living under a rock is one of the big four.

And they want our help, because they did not have access to their own data, they had no way to understand how they were doing with their own data.

And this is one of the largest social media platforms on the planet.

And so what we had to do was looking at third party data sources, external data, publicly available data, anything that we could get our hands on, that would help us put the story together.

Now while it’s not directly tied to your sport house, the reason I share this story is because you may be in a similar situation.

And what we’ve seen is unfortunately, sometimes the larger the company, the more bureaucracy, you run into even getting your hands on your own data question systems on your own data on data that’s not already touched already filter already calculated on getting that raw data.

So I would encourage you to sort of think outside the box and Okay, what else external data can try to get that is available that could start to answer some of those questions.

And so we can make some assumptions about what’s happening.

And then show this is why we need our own data.

That’s why we need access your own data.

The second example is, I used to work with a very large UK company, they’re a weight loss company, and they were trying to break in to the US market.

Now.

The US market is saturated with weight loss companies, you have Weight Watchers, Jenny Craig, you have all kinds of crazy diets.

And they want to break into that market.

And they want to do it in a digital way.

So they wanted to do as they want to do content.

So every week, my team would deliver them a new, here’s your time series forecast on the terms that you want to be known for on the on the places where you should spend money.

These are the hours that you should promote these little ones that you should tune down, this is a full year 52 week plan 365 days we’ve literally handed to you just do this thing.

And you will get great results.

They did not have it, which is really frustrating for a lot of reasons.

But they did not have it because they felt that their guesswork their assumptions, their feelings about their customers were better than the data that we had.

Well, I’ll give you two guesses as to what happened the first, the first one doesn’t count.

They fired us because they didn’t get the good results that they want it but they didn’t follow the plan, they didn’t use the data that we were giving them that said, this is exactly what you should do.

Now, there’s not a whole lot you can do in that situation, you can lead a horse to water you cannot make you put on a bathing suit.

So you can only take a person so far, when you give them the data and say this is the exact point, you’re always going to run into some resistance, which is why again, I encourage you start with a smaller proof of concept.

So people can really see the value and understand this is how this works.

So you actually have a real life case study for us.

This was a client that we worked for their publishing companies content, but we their b2b company.

And we did the content plan that I showed you.

So we took a look at all the terms they want to be known for.

And for one full year, we used that time series forecast.

And we optimize their content, we help them create new fresh content and made sure that it was being published on the right times.

Because when we looked at their results in 2017, we said you know what? These are good, but they could be better.

How can we do better.

So by using a time series forecast, specifically on their content, what we saw was that it drove users up 80% year over year.

Okay, that’s pretty good.

Even better.

Using that time series forecast resulted in a 285% increase in conversion.

And in this case, conversion being someone taking action on the web site, filling a form raising their hand saying I want to work with you, I want to buy something from you.

It also resulted in a 248% increase in revenue.

I encourage you to use a time series forecast, you can achieve results like this, I feel like I’m on an infomercial, I apologize.

The point being is that using the data to help you understand what to do when really takes the guesswork out of it.

And then you can just do the thing and focus on building those deeper customer relationships.

You can focus on those actions, those insights, those strategic plans, take focus off your plate, this is nonsense.

Let them shoot into it for you.

So what do you do next? evaluate your goals.

Figure out your business strategy, figure out is your company ready for something like this? Do a small proof of concept.

Make a plan? What’s the question you’re trying to answer? What are what are the pain points that you’re facing right now, as a team as an individual that a time series forecast couldn’t’ve potentially help you figure out, get some small wins.

Think again to that proof of concept.

What is a very small proof point that I could do to demonstrate this is how this works.

And then you start to iterate and adjust the plan and measure it make sure you measure everything.

Take your baseline metrics before you start, take your metrics afterwards.

That is the most important thing because that’s how you’re going to demonstrate progress and how this pretty good forecast can actually work for you.

And that’s it.

Short and sweet.

Any questions? We have time.

everyone just wants to run to the open bar in serious any questions

Need help with your marketing AI and analytics?

Machine-Generated Transcript

Leave a Reply Cancel reply

Subscribe to our Weekly Newsletter

Pin It on Pinterest