PredictiveEverywhere

Please post your comments here!

Integrate data silos; pick a reliable, valid metric.

For starters, you need to aggregate accurate customer data in order to build statistical models that drive valid predictions.

Recognise success and failure as you begin data crunching. Know when to draw the line in experimentation.

I agree with B.O.B. Break down those data silos and integrate data across all biz units (sales, marketing, call center,...)

Also, accept fast-fail as a best practice -- try things, low hanging-fruit, quick wins, and quick fails. Learn from all.

You also should be building segmentation models according to key features/attributes--income, socioeconomics, education, geography, etc.--that describe customer behavior.

Thanks you @jameskobielus -- segmentation is key, especially hierarchical models http://syntasa.com/t...

Clarify what exactly you'll be trying to predict: churn, upsell, etc.

As in any #datascience project, start with your business questions and goals. (As @jameskobielus has said :)

Yes to @jameskobielus. You need to identify your problem area. Pick your outcome measures to match your biz needs.

Identify the key analytic apps within which those predictions will deliver value--such as marketing campaign optimization, recommendation engines, etc.

I agree with @Sam___Hurley -- know when to stop cutting bait and start fishing. Avoid paralysis of analysis.

I see too many companies simply pick the NPS as the measure of customer loyalty. There are other, better measures to predict other types of loyalty behavior (e.g., stay, upsell).

Thanks @KirkDBorne! An endless experiment never concludes anything...

Time for question number 2 at the top of your screen.

Identify the range of customer data sources--structured plus unstructured (e.g, internal emails, customer feedback and even comments). Define pipeline for consolidating, preparing, sampling, and staging them for downstream analytics

I use score cards with real-time customer data that are constantly updated with real-time input.

Please post your comments here

waiting for your comments

You should look at historical transactional data to identify which customers seem to be tapering down their visits, purchases, etc. Perhaps they've already effectively churned away.

The signal is sometimes in the noise: lots of customer chatter, call-center calls, & logins to their account may indicate "I am not satisfied with this. I'm outta here"

A popular one is that you ask them in a survey (some evidence that "likelihood to churn" is a valid predictor of actual churn). Also, with integration of data silos, you can run variables through a regression model to identify the best predictors.

Identify patterns of behaviours such as number of times logged in, time spent on activity X etc and compare with historical data sets. #PredictiveAnalytics

Interesting bi-modal distribution: either "no calls or transactions" or "too many calls" may be an early warning sign of departure.

Analyze unstructured data sources such as call-center logs and use natural language processing to identify textual evidence of persistent/growing customer dissatisfaction.

Pattern changes in orders, communications and/or shifts in responses to inquiries.

I agree with @jameskobielus -- #NLProc and Unstructured Data are becoming essential biz assets to address this challenge.

Do social-listening analysis to identify patterns of negative commentary, which may be trending across all customer segments or be highly correlated to particulars (e.g., women 18-40, people in eastern Ontario, etc.).

I like @craigbrownphd suggestion -- look for "stationarity changes" in customer signals (orders, communications, responses)

I'm concerned with over-reliance on #NLProc to measure something reliable enough to have predictive power. Sounds like a study is in order (rating scales vs. #NLP).

Social Listening is a big tool, but need to understand difference between opinion, sentiment, & influence

@KirkDBorne Right. "Real-world experiments" such as gauging differential responses to particular offers among particular segments might indicate that some segments more likely to churn than others.

Time for question number 3 .Please look at the top of your screen for question #3 please.

I am a big fan of A/B testing -- i.e., Experimental #DataScience :)

Context is King! incl. geospatial-temporal data (thank you @madiakc for that suggestion)

Please post your replies here

This is the last question for the day and we request everyone on this chat to chime in with their views

Predictive customer analytics results in an algorithm (regression equation) that tells you which variables are important in predicting customer sat/loyalty. Use that algorithm in #CX management efforts to make decisions/build better apps to assist emps.

That is the ultimate golden goal: turning accurate Predictive Models into Automated Triage = next-best action!

this is where real value is created. Define your predictive model, predict next best action and (in real time) prescrib next best action. Actions can be send to any front end system

+1 @bobehayes > yes! Find those most explanatory variables and exploit them!

Automated next-best actions ride on decision automation, journey modeling, behavioral analytics, predictive analysis, rules engines, and real-time streaming. It's the soul of recommendation engines for e-commerce etc.

Last 5 minutes of the chat left, closing remarks everyone?

+1 @Ronald_vanLoon > Going from predictive to prescriptive models is where all businesses should be driving toward.

BTW : Don’t forget to join us for our webcast : Data-driven customer engagement: Retain valuable customers and grow your business with predictive analytics, on Wednesday, May 18, 2016 11:00 AM EDT : Register : http://ibm.co/1WXL2I...

The #DigitialMarketing pot of gold: Cognitive Analytics = the right decision, right action at the right time, in right context, for the right customer

The next optimal action may be to redirect the customer away from engaging with an automated bot toward a human being who can manually come on to serve their needs and make them feel special.

@bobehayes absolutely right! also an automated feedback loop to access the success or failure of models is also necessary. Regular tracking and tweaking of model makes the system more robust and accurate

Pick the right outcome measures (based on your problem statement). There are different types of customer loyalty (recommendation, upsell, cross-sell, churn), so pick your metric based on your company's needs. Outcome measures have different predictors.

@craigbrownphd @Ronald_vanLoon @KirkDBorne @thealikhanafer : Closing remarks? Last 2 minutes left

@bobehayes @jameskobielus : Closing remarks?

+1 @bobehayes Love those metrics! #DataScience should be all about measurement, experimentation, refinement, and finding the next best action!

Thanks everyone for participating!