Cranes Software has been a strategic partner to CNN IBN in analyzing exit and opinion poll data for the recently held Lok Sabha elections. The exercise is an example of the Company’s strong capabilities in the field of predictive analysis. It has also leveraged its deep domain expertise in statistics and graphics, and used the analytically intelligent tools - SYSTAT and SigmaPlot, in providing back end data analysis for the on screen graphics. Dr. Rajeeva Karandikar, a well regarded probabilist and statistician and Executive VP at Cranes and his team has been associated with the news channel for all election forecasts over the past three years.
Talking about how he went about getting it right, Dr. Rajeeva Karandikar, explained: “The opinion polls that predict the composition of Lok Sabha or Vidhan Sabha have attracted a lot of attention over the past 20 years. While these polls provide fodder for debate across media channels; one cannot ignore questions such as ‘how can surveying a small fraction of people give insight into how the country is going to vote?’
We could try discussing the situation and process through a set of examples.
In 2009, the total electorate in India was 67crore and an estimated 32 crore voters exercised their power to vote. Our sample of about 30,000 constituted a minuscule fraction of this total. Is this small sample a fair representation of the total number of voters, and can these people provide accurate insight into how the entire nation is likely to vote?
The answer to this is: a carefully chosen sample using stratification and randomization is most of the time, representative of the population and thus can be used as the basis for a broad judgment about the behavior of the entire country.
For the 2009 Lok Sabha elections, a survey was conducted by Center for Studies in developing societies (CSDS) as National Election Study (NES). CSDS has conducted election surveys for several decades. CSDS provided state wise estimates of vote shares for major parties along with other information to the television network CNN-IBN.
However, the interest of a common man is in seat projections and not in vote share projection. How do we go from votes to seats? One option could be to borrow methods used in the United Kingdom since their political system is close to one followed in India. However, in UK the voting intention is fairly stable from one election to the other (if we defined to be the proportion of voters who voted for a different party this time than they did vote for in the previous election, then in the assessment of experts as well as empirical evidence based on surveys,d is fairly small for UK). On the other hand, in India voting intentions are volatile and are believed to be fairly high. In a nationwide survey done in 1998 by CSDS,d was estimated to be around 30% over a period of four weeks.”
In the current engagement of Cranes Software with CNN-IBN, the role of Cranes and in particular, the role of Rajeeva Karandikar was to project seats in the next parliament, based on the state wise vote shares estimated by CSDS. Dr Karandikar has been an advisor to CSDS on their survey work in the past, including the 1998 survey referenced above.
Continuing his explanation, Dr. Rajeeva Karandikar, said: “A simple mental exercise would convince the reader that in a first past the poll system (where the candidate getting the highest votes in a constituency wins the seat) practiced in India, there is no simple rule to translate vote shares to seat shares. Hypothetically, if there are two parties ‘ABC’ and ‘XYZ’ in an assembly of 100 seats, with ‘ABC’ getting 50.1% votes and ‘XYZ’ getting 49.9% votes, it is possible for ‘ABC’ to win all the seats- this would happen if in every constituency it wins 50.1% votes. On the other hand even ‘XYZ’ with 49.9% votes could win all but one seat- in 99 seats it could have say 50.1% votes and in the 100th seat it could have very few votes (0.3%). This example depicts that to get from predicted vote share to predicted seat share, one needs to know the distribution of votes of each of the parties across the entire state or nation.
The use of raw data is not sufficient, domain knowledge is important. Using the domain knowledge, one needs to model the votes of each major party across each constituency and then estimate the parameters used in the model keeping the objective in mind.
Here the objective is to project seats of major parties in the house (Lok Sabha or Vidhan Sabha). In this context, the following model has been found to be suitable: within a state, the ‘distribution’ of votes for a major party does not change much from one election to the next or to put it differently, the change of votes (also called swing) from one election to the next for a given party is uniform across a state. This model is fairly robust. Using this model, one can estimate the vote shares of major parties in each constituency by using the history data along with measuring statewide vote share of each party. It should be noted that even in election for Lok Sabha, the swing can be modeled as uniform only across a state. Also, for a large state like Uttar Pradesh or Maharashtra, we divide the state in geographical regions and model the swing to be uniform across a region.
Now we need to convert these vote estimate into seat estimates for major parties. We have developed a model for this task. The model takes into account the political realities on one hand and out expertise in mathematical modeling. This technique is called Probabilistic count method and it gives good estimate of likely composition of the parliament.”
This method of converting estimated vote share into seats has been proven to be fairly good and seems to be the best among feasible approaches. The Cranes Software team led by Dr Karandikar has been working with CNN-IBN for the last three years on such psephology related projects. The projections made by the team with the algorithms, etc developed using the IPs of Cranes, for various assembly elections and the recent Lok Sabha elections have proved to be very accurate for most of the elections.
The analytics team at Cranes Software is a pioneer in providing statistical modeling and solutions proof of concept and delivery. The group focuses on target markets of pharmaceuticals, environmental science, social science, telecom, and BFSI. The Company has successfully accumulated a portfolio of proprietary products in all elements of the analytics value creation process. In doing so, Cranes has built a library of reusable components be it statistical graphing engines, computational algorithms, or data integration interfaces. These reusable components help the group of consultants do quick prototypes of customized solutions to clients and provide enterprise class decision support systems. Some products within the portfolio of Analytics domain include SYSTAT, AISN Data Visualization Software Suite, Sigma Products Suite, iCapella, InventXePM and the recently acquired Cubeware and Dunn Solutions Group.
About Cranes Analytics Group
Cranes Analytics, a strategic business unit of Cranes Software International Ltd., is a leading provider of high functionality enabled Statistical Products, BI consulting, Data Analytics, Statistical solutions and software development services for the BFSI, Telecom, Pharma, Healthcare, Life science, Social & Environmental Science industry verticals. The division operates in a very specialized business category and has three distinct streams Scientific Analytics that include statistical and mathematical modeling products for high end scientific applications. Enterprise analytics that combines business and statistical intelligence for enterprise wide business reporting and on-line decision support. Enterprise decision support systems, in which Cranes had recently acquired Cubeware, a mid tiered BI company from Germany that has its own data integrating and information dash boarding technology. The division also has alliances partnerships with Microsoft, Business Objects, BEA, Borland, Net Solution and Craft Silicon.