Categories

Archives

Understanding Numbers for making Decisions (II)

A) Index Numbers: Some times you have numbers of different ranges, however you may want to compare on a single scale. Indexing numbers will help you achieve that. For example average sales of apples  may be 1/2 of average sales of oranges for a given month/geography. If you compare  their trend of sales, you find that they are at same level from their repectives bases. How do you capture that information?

You fix a particular performance  level as base, you can assign a metric as 0 or 100 or of your liking. All other performances are Indexed on base.

This will help you establish high/low movement of metric from base.

Therefore sales of apples and oranges, even though are different may be at same Index point. On the other hand grapes sales may be same as apples but its index may be 1.5 times that of apple.

Another use of index is offering points to sales guys for their performances. Every sales guy on field may have different targets to achieve. However they may get points depending on the ratio of their performance to their base level performance.

A particular state of performance is considered as BASE then the future performances are treated as ratio from base expressed in range of 0 – 100 or 1 – 5 or 1 – 10.

You can square the number before taking ratio. Alternatively take root, mod, absolute, log etc depending on how you want to represent numbers.

B) Conversion Rates / Probability: Whenever there are two levels of metrics one leading to another. Conversion rates are normally used. Most commonly used conversion rate is (Customers acquired / Prospects contacted).

This can be used as performance of channel used for identifying propects. In internet businesses Conversion rates are used for evaluating performance of landing pages, acquisition channels (SEO, SEM, Affiliate), product efficiency, managing advertising and many more.

Conversion Rates can also be used as probabilities of occurence of the event. For all practical scenarios you can use this as a metric to analyze effectiveness of campaigns.

Frequency distribution: In other words it is proportion of segment compared to whole. This can also be used as probaility.

In futre topics, lets talk about how we can use this to calculate lift, ROC etc.

For now it this will help point out Outliers in database, segments that take up large portion of the pie etc. This metric will also help you in segmenting database.

C) Ratios: When you divide one number with the corresponding number of a different metric its a ratio. Conversion rate is also a ratio.

Odds Ratio: This is defined by probability of occurence of an event over probability of non-occurence of that event. In some cases it is (Conversion rate / (1 – Conversion rate ) )

This is mostly used in logistic regression to estimate the probability of occurence from a set of predictor variables.

You could run a predictive model to predict the log of odds ratio as the outcome (dependent variable) and input independent variables to be customer predictors.

Predictive analytics is good at identifying, whether a event will occur or not, rather than when the event will take place.

Further in the series of understanding numbers, we will talk about concept of lift and other metrics.

  • Share/Save/Bookmark
July 30th, 2010 | Category: analytics | Leave a comment

Understanding Numbers for Making Decisions

Every company has data in some form or the other. This data may or may not be used for making decisions. The most important data for any company is invariably the revenue or sales data. Its obvious that this data is important.

The important aspect is how this revenue data is used to identify impact due to other metrics or data. One big mistake people make is looking each of the data independently. This robs of the wonderful insight that seen from impact analysis.

Few of the important ways in which data can be understood are.

A) Combine Relevant metrics. Look at data in relation to other relevant metrics to understand the impact. For example instead of looking at revenue separately and expenditure separately we need to look at data revenue per every unit of expenditure. Similarly for other metrics, you can calculate ARPU, average revenue per user. Likewise average revenue per visitor and metrics that is actionable.

For example looking at telecom data, where average height of students in a class is of little importance. Instead you can look at average height of boys, average height of girls. Similarly average height of boys/girls who are active in sports can be of some value.

Another example is let’s say a company has 17% attrition rate (churn rate), and company has 400 employees with Revenue of $2,800,000.

You can look at this data as average revenue / employee as $7000 and 68 employees resign every year with a total of $476K. A delay of 1 month in hiring will cost you $40K. Look at data in combination and not independently.

B) Secondly looking at numbers at an average level does not make sense. You need to segment the data into relevant categories to make business sense. For example how much sense does it make by looking at average marks of students in 10th std for all subjects.

You need to look at each subject, each division, breakup on metros, rural/urban breakup, break on central/state boards etc.

Similarly looking at any data, segmentation is the basic key to understand it. You need to segment data in all possible ways across n x n dimensions that is possibly available.

C) Next is looking at data from paretos angle. Paretos principle says there exists a 80-20 rule. 80% of the revenue comes from 20% of customers. You can try and apply this paretos principle to as many situations you want. Average numbers may not even make sense in most cases.

You may find that 20% of the students score above 90% marks and rest of them score less than 75% marks, bringing the average down to 80% marks which is an imperfect number to look at.

D) Absolute Numbers & Percentages: Looking at numbers both in terms of absolute numbers as well as percentages is good. For example a distribution across geography may result in higher % in some locations, like for instance let’s say Chennai may contribute to 5% sales, however absolute numbers may be 10 to 15 units in number which may be insignificant. Need to be careful looking at numbers/percentages.

The advantage of the % is it offers the slice of the pie.  Absolute numbers gives us idea of volume of data, Percentages hide the volume.

The article is originally found on http://www.praveenkodur.com/blog

  • Share/Save/Bookmark
July 11th, 2010 | Tags: , | Category: analytics | Leave a comment

KPIs Business Intelligence Dashboards

Business Intelligence is very important for every company, it is the sole heart of decision making that happens in the company. Most companies however spent little time and money in getting the right business intelligence dashboards for the company.

Few of the key things that could hamper their decision making is not having a consistent definition of metrics across the company. Finding cheap and effective way to store business information with minimal inconsistencies is essential. Along with this is required a solution that is scalable for data explosion and variety of future business / analytical needs.

There are lot of BI tools in the market from Top Premium tools to low cost open source tools. Finding the right tool and people for maintenance of the tool and even more people for using the tool is essential for every company.

Analytical needs of the company requires data capturing at granular levels. This transactional data needs to be captured for business intelligence at higher levels. There are lot of new technologies that are currently available such as databases on schemaless Mapreduce architecture or Databases on Cloud computing etc..

These new technologies can make the Implementation of BI much more simpler and easier than before. At the client end however, you need tools for analyzing and manipulating data.

Even web analytics tools like Google Analytics provides tools and features to create your customized reports which can be viewed online or scheduled to your email address. The most important thing however is to identify the few key metrics that matter to your business.

Its common that most business at the start want to look at total pageviews, total visitors as metrics in the dashboard. However everyone knows these metrics do not change on day to day basis. The report ends up showing almost the same data every day.

Identifying key metrics can be difficult and needs feedback from business heads and top management. Lets  take an example of Online Marketing, instead of generic report on total acquistions, total cost, pageviews, clicks, CTR, Revenue, ROI.

You could make different dashboard which contains – Profit from Branded Keywords, Profit from Long tail keywords, Profit from Acquisition keywords. Similarly Bounce Rate of landing page for each of the keyword segments. Another metric could be ARPU (average revenue per user) from each of keyword segments.

Since the action points are at keyword level or banner level. The segmentation is required first at keyword / banner level. You have to segment keywords/ banners based on their value delivered.

In case value is not directly visible as in revenue or profit we need to define proxies to identify the value. Some keywords may attract users who are more engaging on the site. This engagement has to be identified.

Few Laws of BI:

First law of BI is anything that matters is detectable. If a metric does not matter, then you dont care if it is detectable. Second law is anything that is detectable is measurable.

Therefore if a user engagement is important and matters to business then it is detectable. Engagement is different for different sites. For blog may be it is adding comments, for a e-commerce site it is browsing more products, for a classifieds site, it may be number of searches. For a social networking site it is number of pokes, scraps, wall posts etc. You have to define for your engagement metric for your business. Now go back and segment your action points with your engagement metric.

Hope this is useful..Please share your thoughts .. send in comments..

This article is originally found on http://www.praveenkodur.com/blog/

  • Share/Save/Bookmark
June 12th, 2010 | Tags: , , | Category: analytics | Leave a comment

Mathematical Models for Online Business

Mathematical models are developed in every area from atomic energy to bio technology.  Models are generally designed to understand the real world phenomena mathematically/algorithmically. A model is a prototype of reality or an abstraction. Real world contains a lot of assumptions and axioms, despite this we can fit the real world process  into a mathematical model.

Every process, business or company can be considered to contain a list of inputs, transformations / systems (working) and finally the output. Of course all of these operate in a set of external and internal conditions (called environment). This real world phenomena can possibly be represented in mathematical form in terms of flowchart & numbers.

Pictorial and mathematical representation will help in optimization and alignment to the business KPIs. Coming to KPIs, these are set of metrics which measure performance of the business. These are performance metrics that are tracked, measured & monitored on regular basis.

Mathematical models contains an additional loop (feedback loop) in normal flow chart, gets formed from output to system as a feedback of performance. This feedback contains information on what to improve and  other intelligence that is fed back into the system. The feedback can be however not necessarily limited to effectiveness and efficiency of the system. First is the ability of the system to meet the desired output/ objectives. Second is the relationship between input and output, measure of how good is the relationship.

More articles on analytics, online marketing is present on http://www.praveenkodur.com/blog/

Users must understand that all mathematical models will only help support a decision or help solve a problem. At the time of decision making a knowledge worker generally has one or more alternatives to solve the problem. A list of feasible solutions will need to be chalked out from all possible solutions. The best solution or optimum solution to be derived from the list of feasible solutions again based on criteria such as profitability, cost, response etc.

  • Share/Save/Bookmark
August 6th, 2009 | Tags: , , | Category: Internet Marketing, analytics | 6 comments

Marketing Data Mining – Its easy to do badly

Companies are slowly adopting to the data mining as a concept, they are looking for quick and easy solution for their problems. On the other end there are whole host of companies which are offering software solution and tools for data mining.

There is a danger here, the easy use of GUI tools on large amount of data available is tempting and which makes users tempt to use black box methodologies available in the tools to solve their business problems. They mistakingly assume data mining is all about using a tool and running the data on the tool, this can be very hazardous and dangerous as actions are taken on business decisions.

Little knowledge is dangerous while applying powerful models.

Since the underlying logic used for model building in automated tools are unknown (their internal assumptions are unclear). Therefore it is very easy to do it in the wrong way assuming it is the right solution. Such mistakes are still made in companies.

People sometimes might argue that there is always a simpler solution to problems. I am the person who belives in Occam’s Razor, which means simplest explanation is always the best way. This means the interpretation of the model and the final solution should be as simple as possible. However,  method of arriving at solution must be as detail as possible taking every factor, element into consideration. Analysts must not make the mistake of over simplyfying the method and make too many assumptions.

A lot of knowledge workers believe that predictive modeling process can be automated by using tools and statistical software. These software are certainly useful but cannot replace the intellectual  of the analyst.

The processof building the model must be as detailed as possible, it has to tried & tested on various data and measured by various parameters. All possible factors must be considered while building the model.

Therefore Analyst must understand the underlying algorithm , process of arriving score, model design etc.

More on such articles regarding online marketing, data mining, data modeling can be found at http://www.praveenkodur.com/blog

  • Share/Save/Bookmark
August 1st, 2009 | Tags: , , | Category: Google, Web Analytics, analytics | Leave a comment
 
  Older Entries »