7月 202017
 

For many industries, products and features are no longer the most crucial differentiators in the minds of customers.

Take mobile telecommunications, for example. The recent market shift from virtually no unlimited data plans to announcements of unlimited data offerings by every major US wireless carrier in a short span of time left many consumers wondering which plan had the best options for their lifestyle and budget.

As I read comments on a variety of blogs, it was apparent that consumers were confused, too. But by and large, they stated they were making choices based on how well (or how poorly) their current providers treated them, not necessarily the ideal plan.

Get this free report to see why SAS is a leader in real-time interaction management

And that’s why customer experience is the real differentiator today. For a sustained competitive advantage, you will want to improve customer experiences however and whenever you can.

Using real-time data for better customer offers

Take for example some recent work I did with a global telecommunications company.

This company wanted to explore the benefits of making more appropriate offers in real time for their customers. The definition of appropriate here being analytically driven with the use of real-time data. Targeted customers were those who reaching their monthly data cap before the end of their billing cycle. The decision point was the offer to top-up a customer’s data allowance.

The decision was based on:

  • Whether to make the offer or not. The company anticipated a low response rate. Sending an offer via email or SMS might be considered spam by the customer.
  • The size of the offer to be made (i.e., how many gigs of data).
  • The cost of the top-up when compared to the projected overage charges the customer would incur.

The way in which the decision could be made was either:

  1. In real time, trigger the same offer to all customers that reach certain percentage of their data allowance. This method was of limited value because customer data was not readily available to the system that executed this trigger. This led to option 2.
  2. Overnight, analyse the customer data to make offers to those more likely to respond, and ensure that the size and cost of the data top-ups were appropriate (e.g. smaller offers to customers closer to their monthly refresh of data allowance).

The result was an approximate 400 percent increase in response rate when compared to the company’s existing methods.

We showed the company how much better the response would be, if they combined the two approaches to make a customer-driven decision in real time.

Our suggested solution was to:

  • Create models based on using the real-time data.
  • Use SAS® Event Stream Processing to identify when customers were running out of data, and trigger an action.
  • Use SAS® Real-Time Decision Manager to run the models in real time and use the output of the models to decide on the most appropriate offer for the customer.

Real-time decisions need to be accurate and effective

The result was an approximate 400 percent increase in response rate when compared to the company’s existing methods. This improves customer satisfaction (more timely and appropriate offers, and not running out of data at an inconvenient time) reduced cost (fewer messages are sent) and increased profit (greater offer acceptance and improved margins).

This demonstrates that there can be huge value in real-time information in the analytics life cycle by creating or improving models and using those decision models to improve the customer experience.

The challenge that companies face is determining which decisions will be most affected by real-time data, and what type of real-time data will be most predictive. But it is not always the case that you need to test (or invest significant resources) to identify these two items because there may be natural tests in the data already.

Say a customer applies for financial services product such as an overdraft, a loan or a credit-limit increase. The customer application is the target variable, and you can use the data (transactions, balances, headroom, future payments, time to salary payment, etc.) as observation variables – taking care to ensure that you capture the timing of these, too – and then identify which of these is predictive of approval or denial. Some of this data will be easier to get in real time than others and this analysis can focus the effort onto the ones that are accurate and effective in real time.

Editor's note: This is the first in a series of posts that offer real-world examples of how to best use analytics to meet your marketing needs. The series will cover several industries including telecommunications, banking and retail.

Real-time decisioning offers unmatched customer experiences was published on Customer Intelligence Blog.

7月 192017
 

If you’re familiar with the SAS story, you know that we have deep roots in academia. But our CEO Jim Goodnight has always known that roots aren’t enough to achieve incremental growth. You also have to plant seeds. Analytics alone can’t drive change, help businesses succeed and make the world [...]

Planting seeds through education was published on SAS Voices by Fritz Lehman

7月 192017
 

A lot of my friends seem to be getting married these days. Which got me thinking about wedding parties. Which then got me wondering what songs DJs do/don't play at weddings these days. And what was the outcome of my meandering thoughts ... a fun & interesting graph, of course! It [...]

The post Songs most frequently banned at weddings! appeared first on SAS Learning Post.

7月 192017
 

Skewness is a measure of the asymmetry of a univariate distribution. I have previously shown how to compute the skewness for data distributions in SAS. The previous article computes Pearson's definition of skewness, which is based on the standardized third central moment of the data.

Moment-based statistics are sensitive to extreme outliers. A single extreme observation can radically change the mean, standard deviation, and skewness of data. It is not surprising, therefore, that there are alternative definitions of skewness. One robust definition of skewness that is intuitive and easy to compute is a quantile definition, which is also known as the Bowley skewness or Galton skewness.

A quantile definition of skewness

The quantile definition of skewness uses Q1 (the lower quartile value), Q2 (the median value), and Q3 (the upper quartile value). You can measure skewness as the difference between the lengths of the upper quartile (Q3-Q2) and the lower quartile (Q2-Q1), normalized by the length of the interquartile range (Q3-Q1). In symbols, the quantile skewness γQ is

Definition of quantile skewness (Bowley skewness)

You can visualize this definition by using the figure to the right. Figure that shows the relevant lengths used to define the quantile skewness (Bowley skewness) For a symmetric distribution, the quantile skewness is 0 because the length Q3-Q2 is equal to the length Q2-Q1. If the right length (Q3-Q2) is larger than the left length (Q2-Q1), then the quantile skewness is positive. If the left length is larger, then the quantile skewness is negative. For the extreme cases when Q1=Q2 or Q2=Q3, the quantile skewness is ±1. Consequently, whereas the Pearson skewness can be any real value, the quantile skewness is bounded in the interval [-1, 1]. The quantile skewness is not defined if Q1=Q3, just as the Pearson skewness is not defined when the variance of the data is 0.

There is an intuitive interpretation for the quantile skewness formula. Recall that the relative difference between two quantities R and L can be defined as their difference divided by their average value. In symbols, RelDiff = (R - L) / ((R+L)/2). If you choose R to be the length Q3-Q2 and L to be the length Q2-Q1, then quantile skewness is half the relative difference between the lengths.

Compute the quantile skewness in SAS

It is instructive to simulate some skewed data and compute the two measures of skewness. The following SAS/IML statements simulate 1000 observations from a Gamma(a=4) distribution. The Pearson skewness of a Gamma(a) distribution is 2/sqrt(a), so the Pearson skewness for a Gamma(4) distribution is 1. For a large sample, the sample skewness should be close to the theoretical value. The QNTL call computes the quantiles of a sample.

/* compute the quantile skewness for data */
proc iml;
call randseed(12345);
x = j(1000, 1);
call randgen(x, "Gamma", 4);
 
skewPearson = skewness(x);           /* Pearson skewness */
call qntl(q, x, {0.25 0.5 0.75});    /* sample quartiles */
skewQuantile = (q[3] -2*q[2] + q[1]) / (q[3] - q[1]);
print skewPearson skewQuantile;
The Pearson and Bowley skewness statistics for skewed data

For this sample, the Pearson skewness is 1.03 and the quantile skewness is 0.174. If you generate a different random sample from the same Gamma(4) distribution, the statistics will change slightly.

Relationship between quantile skewness and Pearson skewness

In general, there is no simple relationship between quantile skewness and Pearson skewness for a data distribution. (This is not surprising: there is also no simple relationship between a median and a mean, nor between the interquartile range and the standard deviation.) Nevertheless, it is interesting to compare the Pearson skewness to the quantile skewness for a particular probability distribution.

For many probability distributions, the Pearson skewness is a function of the parameters of the distribution. To compute the quantile skewness for a probability distribution, you can use the quantiles for the distribution. The following SAS/IML statements compute the skewness for the Gamma(a) distribution for varying values of a.

/* For Gamma(a), the Pearson skewness is skewP = 2 / sqrt(a).  
   Use the QUANTILE function to compute the quantile skewness for the distribution. */
skewP = do(0.02, 10, 0.02);                  /* Pearson skewness for distribution */
a = 4 / skewP##2;        /* invert skewness formula for the Gamma(a) distribution */
skewQ = j(1, ncol(skewP));                   /* allocate vector for results       */
do i = 1 to ncol(skewP);
   Q1 = quantile("Gamma", 0.25, a[i]);
   Q2 = quantile("Gamma", 0.50, a[i]);
   Q3 = quantile("Gamma", 0.75, a[i]);
   skewQ[i] = (Q3 -2*Q2 + Q1) / (Q3 - Q1);  /* quantile skewness for distribution */
end;
 
title "Pearson vs. Quantile Skewness";
title2 "Gamma(a) Distributions";
call series(skewP, skewQ) grid={x y} label={"Pearson Skewness" "Quantile Skewness"};
Pearson skewness versus quantile skewness for the Gamma distribution

The graph shows a nonlinear relationship between the two skewness measures. This graph is for the Gamma distribution; other distributions would have a different shape. If a distribution has a parameter value for which the distribution is symmetric, then the graph will go through the point (0,0). For highly skewed distributions, the quantile skewness will approach ±1 as the Pearson skewness approaches ±∞.

Alternative quantile definitions

Several researchers have noted that there is nothing special about using the first and third quartiles to measure skewness. An alternative formula (sometimes called Kelly's coefficient of skewness) is to use deciles: γKelly = ((P90 - P50) - (P50 - P10)) / (P90 - P10). Hinkley (1975) considered the q_th and (1-q)_th quantiles for arbitrary values of q.

Conclusions

The quantile definition of skewness is easy to compute. In fact, you can compute the statistic by hand without a calculator for small data sets. Consequently, the quantile definition provides an easy way to quickly estimate the skewness of data. Since the definition uses only quantiles, the quantile skewness is robust to extreme outliers.

At the same time, the Bowley-Galton quantile definition has several disadvantages. It uses only the central 50% of the data to estimate the skewness. Two different data sets that have the same quartile statistics will have the same quantile skewness, regardless of the shape of the tails of the distribution. And, as mentioned previously, the use of the 25th and 75th percentiles are somewhat arbitrary.

Although the Pearson skewness is widely used in the statistical community, it is worth mentioning that the quantile definition is ideal for use with a box-and-whisker plot. The Q1, Q2, and Q2 quartiles are part of every box plot. Therefore you can visually estimate the quantile skewness as the relative difference between the lengths of the upper and lower boxes.

The post A quantile definition for skewness appeared first on The DO Loop.

7月 182017
 

I clearly remember the morning last December when I interviewed for SAS. I was visiting my brother in Seattle over winter break, so I was interviewed over Skype by Brandon and Jason (my current manager and mentor respectively). Near the end of the interview, Brandon asked me, “Where do you [...]

The post Sustainability at SAS appeared first on SAS Analytics U Blog.

7月 182017
 

Nowadays, whether you write SAS programs or use point-and-click methods to get results, you have choices for how you access SAS. Currently, when you open Base SAS most people get the traditional SAS windowing environment (aka Display Manager) as their interface. But it doesn’t have to be that way. If [...]

The post Organize your work with SAS® Enterprise Guide® Projects appeared first on SAS Learning Post.

7月 182017
 

In part one of this series, Clark Twiddy, Chief Administrative Officer of Twiddy & Company, shared some best practices from the first of three phases of Twiddy’s journey to becoming a data-driven SMB. This post focuses on phases two and three of their journey. Phase two is about action. Now [...]

How to be a data-driven SMB: Part 2 of Twiddy’s Tale was published on SAS Voices by Analise Polsky

7月 182017
 

Datasets are rarely ready for analysis, and one of the most prevalent problems is missing data. This post is the first in a short series focusing on how to think about missingness, how JMP13 can help us determine the scope of missing data in a given table, and how to [...]

The post How severe is your missing data problem? appeared first on SAS Learning Post.

7月 182017
 

St. Louis Union Station welcomed its first passenger train on Sept. 2, 1894 at 1:45 pm and became one of the largest and busiest passenger rail terminals in the world. Back in those days, the North American railroads widely used a system called Timetable and Train Order Operation to establish [...]

The post Finding important predictors: Using your data to explain what’s going on appeared first on SAS Learning Post.