Week 4 - Algorithmic Pricing

Week 4 – Algorithmic Pricing Standard Pricing: The firm’s main goal is assumed to be maximising profits, marginal cost exists. This also assumes that the firm knows the demand line, and what demand that firm is facing. But, the firm doesn’t know what demand they will have in the future, this shows the relationship between price and prediction. The firm sets the price at where marginal cost equals marginal revenue. Demand is, therefore, estimated from data – how can you set a price whereby its low enough to attract consumers but high enough to maintain a high market share. What techniques were firms using in the past when there was no data on demand or computational capability? - - Market price as a benchmark/competition-based pricing: large number of firms, no firm has an impact on the price of the market, so new firms set their prices to the average price of the market. Mark-up pricing: selling price = cost of goods sold + (mark-up x cost of goods sold) These are both uniform pricing techniques – everyone in the market is charged the same. Mark-up pricing makes pricing extremely easy, but depends on prediction. Firms used to use human estimates. - Gross Margin Pricing: Selling Price = Costs of Goods Sold / (1 – GM) For both this and mark-up pricing, how do you choose the ideal mark-up/gross margin? Predict the future or use past data (but this also predicts that the market will act the same way in the future). - Proportional Pricing: the price is set equal to the base price and multiplied by a constant factor (K) times the ratio between supply and demand. Selling price = pricebase x k x (d/s) There are many unknowns here – how should the base price be set? How do you choose k? how do you systematically forecast supply and demand? All of these are prediction tasks. Uniform pricing, or a given price you start with, may not necessarily maximize profit. Non Standard/Non Uniform Pricing – the firm charges different consumers different prices for the same product. Price discrimination is charging different prices to different consumers for the same good to achieve higher profits. - First degree/perfect price discrimination The firm charges different consumers based on his/her willingness to pay, this changes consumer surplus into producer surplus, and raises the profits. - Second degree price discrimination The firm charges different groups of customers different pricing e.g. peak/off-peak travelling. - Third degree price discrimination Different groups have different willingness to pay, the firm charges each group different prices e.g. discounted bus rides for over 65’s or Spotify Premium 50% off for students. Perfect price discrimination has been almost infeasible as firms did not have information on willingness to pay at the individual level. Theoretically, firms could not predict a person’s willingness to pay, the challenge was prediction. Personalised pricing: - Yellow dot: price Green line: marginal cost An example of personalised pricing is clubcard prices for card members of Tesco. Uber: Surge Pricing - Uber operates in a market with large fluctuations in demand and variable supply of driverpatterns. App openings are a good representation of those who are in the market for Uber’s services and thus provide a nice measure of demand. The algorithm that employs a “surge pricing” approach assigns a simple multiplier that multiplies the standard fare in order to derive the “surged” fare. Surge pricing draws on the assumption that price elasticity of demand, willingness to pay differs among passengers. During surges, Uber can see which consumers stay in the market (i.e. don’t use the app) so are more price elastic than other consumers. This data could then be sold to other retailers, Amazon for example. Digital technologies allow collecting highly informative information on willingness to pay. Markets with Heterogeneous Goods - - Trading is infrequent and asset examples include properties or collectibles. Valuations are tightly but complex, linked to a wide range of characteristics, which are not necessarily easy to quantify. Human eyes and experience have therefore been deemed indispensable in valuing houses or paintings. Auction is seen as an effective mechanism for price discovery. How far may unstructured data capture tacit knowledge/human judgements underpinning sophisticated pricing decisions? Art auctions are typically organised as “English” – ascending-bid, open outcry. Prior to the auction, the auction house experts publicly share a low and high estimate for each lot. They base price estimates on quality, condition, rarity, etc. Auctions involve reserve prices – the item won’t sell if it doesn’t reach the reserve price. - Images can be used as a starting point as unstructured data for machine learning. The machine learning algorithm successfully predict/generate the long right tail in the distribution of art auction prices – with certain works but certain artists fetching very high prices. The predictions of the machine learning algorithms are about 15% less accurate than those generated by the experts. This inaccuracy could be because auction house experts might have access to more information about the artwork’s quality and history or auction house estimates might affect bids. Real estate market has similar problems – real estate agents value properties using their experience in the market. Glaeser et al (2018) – had data on the majority of home sales in Boston from 2003-15. The algorithm had access to no. of bedrooms, land space etc as well as images. The accuracy error was between 7%-15% for pricing of the properties.