In chess, a computer move is one that is only made by a computer but not by a human. One of the most famous examples is the 44th move in the first game of the 1997 match between the then World Champion Garry Kasparov and Deep Blue, IBM’s chess machine, when the computer played an apparently random move (Rd1) before resigning. Later, deeper analysis revealed that the move would have postponed the much later check mate. It was reported that when Kasparov learned of this, he was so deeply shocked by Deep Blue’s capabilities that he resigned in game 2 in a drawn position.
Humans don’t make computer moves, because the benefits materialize much later and are not immediately apparent to human heuristics such as “connect your rooks and place them on open files.” As in many situations of life, these heuristics help human players understand a position. They don’t necessarily lead to the best results, but reliably to decent ones. Computer moves differ substantially in their characteristics from human moves. For example, AI chess engines like Alpha Zero sacrifice material much more often than humans.
In pricing, we are faced with a similar situation. To optimize prices, pricing managers need to cope with substantial complexity and most resort to a collection of pricing heuristics. For example, one such heuristic states that to price a good-better-best portfolio, you need a low entry price to attract customers, you need a high price for the best option to increase customers’ willingness-to-pay, and then a right-priced middle product that the customer is supposed to buy.
At buynomics, we are building a pricing machine to optimize offer structures and prices. Over the past months, we have conducted several A/B tests of the prices set by the buynomics pricing machine against human-made prices. The buynomics pricing machine optimizes prices using Virtual Customers that make purchase decisions like their real-life counterparts – and predict very precisely how real customers would react to price changes. Across the A/B test, the machine prices have significantly outperformed the human prices. Here, we want to share two observations of how machine prices differ from human prices.
To illustrate the comparison, we consider a café that offers three sizes of coffee. As a proxy for human prices, we looked around the cafés in the neighborhood outside of our office and found a typical price structure like this:
Small (0.2l): €1.99
Medium (0.3l): €2.99
Large (0.5l): €3.99
The logic behind these prices is (as above) simple. The low entry price is to attract customers and the high top-price is to make the medium price look reasonable. Then, most customers will go for the medium coffee. We modeled the purchase decisions of 1,000 Virtual Customers (Figure 1), based on actual preferences in similar situations. When offered the coffee portfolio, 367 out of the 640 who bought a coffee, went for the medium size (57%) – probably confirming to the café owner that his pricing strategy worked.
Observation 1: The pricing machine reduces the price spread in a portfolio
Compared to human prices, pricing machines reduce the price range in a portfolio. Here, the price of the large coffee is reduced from €3.99 to €3.39 (Figure 2). This moves a large share of customers to the large coffee, who would otherwise have chosen the small or medium size. In sum, this leads to a profit increase of about 6% (from €1,569 to €1,664), that comes mainly from the higher absolute € margin of the large coffee (with lower % margin!).
Reducing the price spread in a portfolio requires either increasing the entry price or reducing the top price. Both are risky – unless you can precisely anticipate the volume reaction to price changes (here, -15% from €3.99 to €3.39). This is very difficult with a back of the envelope price elasticity calculation, but easy for a pricing machine.
Observation 2: The pricing machine identifies areas of very large price elasticity and exploits these situations
Most pricing professionals consider a price elasticity below -4 very high and would not change a price if it required an elasticity below -6 to be profitable. However, in portfolios with close alternatives, price elasticities can be very high. Figure 3 shows the price elasticity of the large coffee in the range between €3 and €5. At the starting price of €3.99, the price elasticity (for a price increase) is -8. The pricing machine can work with such high price elasticities and use them to optimize the portfolio offer.
Note that, actual price reactions can be even more extreme. In a recent pilot, our pricing machine reduced one price in a portfolio by 36%, and this price change increased sales (as predicted) by a factor of 25! That is, not +25%, but +2400%, which puts the actual price elasticity at about -67.
These are exciting times to be in pricing! As the two observations indicate, machines have a lot to contribute – and they will make many established heuristics obsolete and companies more profitable. We are just getting started, and we will share further observations as we move forward.
For now, however, we close with a note of caution. Alpha Zero, the AI based chess engine, sacrifices material in exchange for a positional advantage more often than humans. However, that does not mean, that simply sacrificing material will make you a stronger player. In pricing, simply applying the observations will not work. For example, simply reducing the price spread in a portfolio will not necessarily increase profits.
Want to know more?