Eric Siegel answers eight questions about predictive analytics

[@dianelang9473]

Great video.. make more .. more .. please.

[@holidayavecrobin]

nice video really informative..

[@mattl5055]

Another problem with Predictive Analytics when applied to business is meeting the statistical assumptions. How could you meet the normality assumption with business data? Most business data is very skewed. You could use non-parametric methods but that is weak.

[@jagusiff]

If you aren't making Predictions and determining when you are right or wrong you have no means of Learning; and, therefore, be able to make better Business Decisions in the future.

[@simeon24]

*cough* prism, riot, perfect citizen, private interest *cough*

[@mattl5055]

Predictive Analytics and 'Big Data' is simply a business buzzword of this decade, you will probably never hear about this again in a couple of years because companies will find out that it does not work. It does not work in business because you cannot obtain data for all of the independent factors that has a casual relationship with the dependent factor, for example if your dependent is sales, how are you going to obtain data for factors such as brand equity, word of mouth sales, creative etc that might have a casual relationship with sales? What will happen when you do predictive analysis is an overcompensation for the independent factors that you have included in your model (see Omitted Variable Bias). Also, with business data, most of the independent factors you believe will have a casual relationship with the dependent factor will likely to show no relationship when you have conducted a sig test. If say the relationship between the factors are strong, how can you be sure that there is no confounding factor involved? Another fault with predictive is the bias modelling software they use (like eviews) which transforms data in a way that it fits into the model. It is like fitting a square into a circle. Because these software do not look at whether these factors should even be in a model, it is up to the user to determine what goes into the model, thereby there would be a possibility of bias. Business people that believes this should read up on 'Cause and Effect', Confounding factors, Omitted Variable Bias, Statistical Errors, Multicollinearity etc

[@afterhourssupport6296]

you couldn't be further from the truth. I have built many models and one good example for an industry that thrives on these models is the debt collection market, I have built many models that predict probability to pay a debt. As for the possibility of bias in a model. Not sure where that comes from. For example, building a predictive model for probability to pay, why would a programmer be bias. If they have a debt with the company it will still be there. You are not able to pick one person out of the list and say 0 probability because you know them as that is not how the models work. Also if you are saying the issue of bias is based on the programmer just trying to make the model look good to get paid then that isn't an option either. Correlation will be out and it simply wouldn't work. As for software like eviews, a predictive model is not a software package it is a program/script eviews may claim they have the ability to generate a predictive model however it simply will not work. Using software like SAS to write the program/predictive model is how it is done (SAS, R and other software). And a predictive model that is working well for one company may not work as well for another. Take debt collection for an example: Build a model for one company and it works well based on their datasets, and sure you could try sell that same model (modify the variables to suit) to another collection company however it won't work, why? Because each collection company has different types of debt in different volumes. The big agencies will have similar debt however they will normally focus on different types. For example, one may have more car loan debt while another has more credit card debts, two very different models and outcomes. Saying that this is all a trend doesn't make sense. Some countries are slow to pick up on this area however it will not fade out. It has been around for years, it is just more people are talking about it mainly due to the advancements in technology that make this type of work easier, hens more people being exposed to it and seeing it as a "NEW and Wounderful" idea. Every large bank and corporation use predictive models, the bank you are with will most likely (if anything like the banks here) will know as soon as you walk in the door if you are suitable to lend money to and what amount as they have these models running all the time to sort out who to push credit cards to and loans. As for data, you ask where would you get the data for brands, word of mouth etc. very easy. Google sell data and they have most of what you need. What you are searching for, why do you think they ask you to sign in to google search and even if you don't it is then related to your IP address. Go to facebook and look at all the ads on there. Start doing some searches (heck even youtube do it) then you will start to see those ads changing to things you have searched. This data is easy to get. As for other data, obtaining datasets with a persons name, address, phone, date of birth, occupation, relatives and so on is also easy to get. From a business point of view the predictive models will be based on their own data anyways so they have what they need unless they want to add geo factors to it. So a trend? No Issues with Bias? No Do they work? Yes, I have built many Can you get one as an "out of the box" solution? No Should business people read up on them more, yep. They will increase revenue, one model in the debt collection sector increased revenue collections by just under $6 million per month. That is an increase that was a direct result of the predictive model I built, nothing changes, no new debt was purchased and the same files and staff working the accounts. The only difference was staff worked the files based on the results of the model. So $6 million per month, given that not all companies are going to see that size increase some will see higher others lower it just comes down to the number of files/customers they have and so on. Matt L: Message me if you want more info, I am passionate about this and have had many clients say to me at the first meeting "Does this really work" and "How do we know the results will be based on your work". There are a lot of people that simply have not read up enough on the area and are not sure, very understandable.

[@mattl5055]

Glen D Lets just say you are modelling marketing effectiveness on sales, using econometrics. And lets say...... research shows that consumers consider price, product quality, convenience and word of mouth when deciding to purchase a product. Price you can obtain from a company's internal data, but how can you obtain the other three important factors? Thats what I was referring to when i mentioned the problem of Omitted Variable Bias.... where you are missing important factors, which will have to be compensated with the transformation of the marketing factors you are able to obtain from company internal or external sources. The result will be distorted, the effectiveness of the factors you have in the model will be magnified over the factors you do not have in the model. Another significant issue with predictive analytics is how to you account for competitors? How do you determine if they have a negative or positive affect on sales?

[@mattl5055]

Matt L The only way to find out is to do some good old market research. Predictive Analytics/Big Data is limited to only showing you 'what' is happening, but to understand 'why' you have to ask the consumer directly via market research otherwise you are just speculating

[@viveksoni8142]

I think it wouldn't be that hard to get that data too. Preference for convenience can probably be measured by the distance of the store from the buyer' house (credit/debit card data, when swiped for purchases). So what proximity did the buyers lie in with regards to the store's location. Were the majority of the buyers local (prefer convenience) or did they drive there from places afar (prefer other factors, say the infrastructure of the store or the availability of the products). As for Word of mouth, you can't of course figure that out on your own, but a simple post purchase follow up using SMS, Email or a Call offering a small incentive in return will get you from where did these buyers hear about the store or who influenced their purchase? Big Data is a thing, because you need BIG amounts to data to conduct a meaningful research. The more the merrier. I hope I made sense.

[@ionlesan590]

Ever heard of A/B testing?