Data doesn’t always contain information

The proliferation of computers has led to an even greater proliferation of data. Home computers store terabytes of it. Large businesses store even more. Consultants and software vendors like to remind us that data isn’t information unless it’s organized for the purpose of doing something useful, and they make a good point. Many businesses have considerable data at their disposal, but they haven’t transformed that data into information that will help them make better business decisions.

It’s important to remember, however, that not all data provides useful information. Failing to recognize this can lead to wasted effort and even jeopardize the adoption of analytics within an organization.

An example I’ve run into with surprising frequency involves the use of past sales data to set future prices. In many instances sales data contains useful information for price setting . . . but not always.

To understand how sales data can prove inadequate, consider the data shown in Figure 1 collected by a street vendor selling cashews. At a price of $4 the vendor sold 200 bags, but at a price of $5 he sold 300. How could this be? Does demand for cashews go up when they sell for a higher price?

Figure 1

The problem is that time hasn’t been accounted for. The vendor’s observations were taken over a period of three weeks, with the price set at $4 for one week and $5 for two. Figure 2 shows average sales per week as a function of price. It’s now logically consistent to estimate the demand curve based on the observations. Without knowing how long the different prices were posted, the sales data doesn’t contain sufficient information to provide guidance in price setting. Even worse, the vendor could be led to a wrong conclusion. The price of $4 yields weekly revenues of $800; the price of $5 yields only $750.

Figure 2

The importance of incorporating time is so obvious that the example might seem like trickery. But the problem is very real though often less obvious in other contexts.

Consider the sale of plywood by a building products distributor. Contractors needing plywood call to request a price for a stated quantity, and sales agents for the distributor look at inventory, market prices and supply a quote. The contractor then makes a purchase or hangs up the phone. Figure 3 depicts a typical collection of historical sales transactions for a distributor. Distributors have computers brimming with such data and might reasonably ask how they can use it to improve their pricing.

Figure 3

Faced with the data in Figure 3, what can be said about setting future prices? One thing that cannot be determined is a demand curve. Even though the figure plots demand against price, attempting to fit a demand curve to the data is as erroneous as trying to fit a demand curve to the data in Figure 2.

The problem stems from the fact that no price is posted to begin with, much less a length of time a price was posted. Most of the examples we run into as consumers are of the posted price variety: food in the supermarket, apparel in the store, books online. When goods are sold in this manner, logically consistent ways exist to estimate demand at different price levels, just as the cashew vendor did.

However, many transactions — especially transactions between businesses — aren’t conducted on the basis of a posted price. With no notion of a posted price for a fixed period of time, fitting a demand curve to the data simply makes no sense. Unfortunately, I’ve witnessed this and similar mistakes on more than one occasion.

Can anything useful be inferred from the data in an effort to set better prices? Supplied only with data about price and quantity sold, not much. Even vast stores of seemingly useful data don’t necessarily contain information that can serve the desired purpose. ‚”Water, water everywhere,” wrote Samuel Taylor Coleridge, ‚”[but not a] drop to drink.”

This isn’t to say the data in Figure 3 isn’t useful; just that by itself it doesn’t supply useful information. If additional data is gathered, such as quotes that didn’t result in a sale and prevailing market prices at the time of a quote, then models can be developed to make future pricing recommendations — though not through the estimation of a traditional demand curve. Or it may be possible to use related data already being gathered. For example, in addition to price and quantity, transaction data normally includes information on the customer who made a purchase. Past history can be mined to determine which customers buy most often or in the largest quantity, and this information can be used in preparing future quotes.

Analytics consists of a powerful set of tools with which to make better decisions based on data. But to do so, analytics must be applied to data adequate for the task at hand. Sometimes data that seems to contain useful information doesn’t. In such cases it’s necessary to augment the data or look for related data that could prove useful. Above all, it’s important not to make inferences that can’t logically be made from a given set of data. Nothing’s worse than devoting time and effort to an analytics project only to arrive at conclusions that are bad for business.