1 / 0

Learning about Consumers and Markets using Internet Data

Learning about Consumers and Markets using Internet Data. Liran Einav (Stanford and NBER) Based on research with Jon Levin and Neel Sundaresan (and Theresa Kuchler , Chiara Farronato , and Dan Knoepfle ) Big Data in Finance and Insurance 7 th Financial Risks International Forum

rue
Download Presentation

Learning about Consumers and Markets using Internet Data

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Learning about Consumers and Markets using Internet Data

    LiranEinav (Stanford and NBER) Based on research with Jon Levin and Neel Sundaresan (and Theresa Kuchler, Chiara Farronato, and Dan Knoepfle) Big Data in Finance and Insurance 7th Financial Risks International Forum Paris, March 21, 2014
  2. General Motivation New data and “big data” lead to new opportunities for empirical research in economics Much (but not all) of these data opportunities come from online markets, which is the context of today’s talk A naïve view: these granular and massive amounts of data would substitute for economic theory or econometric research design More likely (we think) is the opposite: With one fact, the added value of theory may be questionable, but with thousands of facts, some theory is essential to organize these fact Unlike Netflix or Amazon recommendation systems, or credit card fraud detection, most economic research questions require out-of-sample predictions, so predictive statements need to rely on econometric research designs (although machine learning can certainly help!)
  3. Today’s talk Our project relies on data/research collaboration with eBay to study online commerce See what we can learn about consumer and firm behavior, competition, market structure, and platform design in a large online marketplace Think about empirical strategies that might be useful for taking advantage of internet data Main challenge: unless one focuses very narrowly on a specific product or segment, the “product space” is huge and unstructured Plan for today: describe some of the work we’ve done … mostly by providing illustrative examples of scalable empirical strategies and of some selected results
  4. Learning from Seller Experiments Internet has reduced dramatically the cost of varying displays, prices, and many other parameters, and capturing the results From a research perspective, this massive variation is Great! … lots of “good” variation due to active and passive experimentation Challenging … difficult to isolate specific effects, and customization of products and services raises concerns about selection and endogeneity Idea: exploit “seller experiments” --- frequent practice of sellers listing items multiple times, while varying sales mechanism, pricing, etc. Relative to (a) field experiments and (b) regressions on scraped data, maintains internal validity but takes advantage of scale and heterogeneity of internet markets
  5. Typical eBay Listings
  6. Typical eBay “Seller Experiment”
  7. Data Construction Start with universe of eBay listings (≈ 1 billion / year) Group listings with identical seller and item title, for each year Majority of eBay listings have at least one match; “standard” items Think of each matched set of listings as a “seller experiment” Each has same seller/item, may have variation in offer terms Typically many thousands of experiments (or more) with variation in a given parameter (sale format, auction start price, shipping fees, etc.), but each one is relatively small – ~30 listings on average Basic strategy: run fixed effect regressions exploiting within-experiment variation in prices, fees, displays, other parameters Some potential issues: Focus on “normalized” prices to make items comparable Use finer definitions of “seller experiments” to deal with potential selection and endogeneity concerns The feasibility of these type of checks are yet another advantage of big data
  8. Effect of Start Price on Probability Of Sale
  9. Effect of Start Price on Sale Price
  10. Auction Demand Curves Each start price (s/v) gives rises to a different (Q,P) pair
  11. Effect of Start Price on Probability of Sale
  12. Effect of Start Price (by Category)
  13. Effect of Start Price (by Category) Q/P=-2 Q/P=-3 Q/P=-4 Q/P=-6 Q/P=-10
  14. Consumer Search and Auction Prices Low search costs on internet should reduce price variability Large literature, mostly on price comparison engines, contradicts (caveats: usually posted not transacted prices, sellers differ, etc.) If search costs persist, are they “excessive”? Lee & Malmendier (2010, AER): case on eBay in 2004 where a board game was available at a posted price of $129.95, and sold for more in the majority of parallel auctions …. auction sellers “fish for fools” Look at distributionof auction prices relative to posted price of the identical item sold by the identical seller Data is based on roughly 20k items from 2003 (750k auctions), then 80k items from 2009 (4m auctions)
  15. Distribution of Auction Prices (2003) Auction prices are dispersed, but on average not much lower than posted prices.
  16. Distribution of Auction Prices (2009) Auction prices in 2009 were significantly lower than posted prices!
  17. Auction Discount from Equivalent Posted Price
  18. The Decline of Online Auctions Auction share of eBay active listings and gross revenue Einav, Farronato, Levin and Sundaresan (2012): “Sales Mechanisms in Online Markets: What Happened to Internet Auctions?”
  19. Demand Curves -- 2003
  20. Demand Curves -- 2009
  21. Sales Taxes and Consumer Demand Debate about the collection of sales tax for e-commerce In the US, sellers must collect sales tax only if seller has “nexus” in the buyer’s state; otherwise buyer obligated to pay equivalent “use tax”, but most do not ... Does it affect consumer purchasing? One view says tax advantage very important; alternative view is consumers don’t pay much attention to taxes Answer may affect the impact of policy on offline-online substitution and on the geography of e-commerce “Seller experiments” strategy can get at a related aspect (effect of shipping fees), but taxes are not set (and cannot be varied) by seller, so need an alternative
  22. Amazon thinks tax are important …
  23. Estimating Tax Sensitivity Using “Surprises” eBay search results page (for “hat”): Key point: no info about seller, and in particular about his location.
  24. “Tax Surprises” Approach (cont.) eBay item page: So a NY buyer would get hit with an extra 8.875% in sales taxes, and we can see whether he/she is less likely to purchase (and perhaps more likely to keep searching and purchase something else) Estimate “tax sensitivity” by comparing purchase rates => non-trivial effect, but not nearly as high as for retail price change
  25. Item-Level Estimates of Tax Sensitivity Multiply estimate by 0.79 (one minus purchase rate) to get elasticities: -.93, -1.68, -1.49. Note: column (1) compares in-state and cross-state purchase rates; (2) and (3) compare in-state purchase preference of different states.
  26. Item-Level Substitution Run (conditional) logit regressions of whether purchase something else(after viewing selected item) on key covariates and item fixed effects.
  27. Back-of-the envelope calculations Combine the above with more aggregate results to arrive at some potentially policy-relevant estimates One should take these rough calculations with a large grain of salt. State lowers sales tax by 1% (as e.g. California did in summer 2011) Lowers online purchases by 1.5-2.0% But raises online purchases from home-state sellers by 3-4% Broad switch to taxation of internet purchases Population weighted average tax rate is 7.25% Applying that to all online purchases => 7.25% *1.76% = 12.8% fall. With such a large change, might want to consider pass-through decisions: if sellers absorbed half, purchases would fall by 6.4% and margins by 3.62%, if current margins are 30%, a substantial change
  28. Final thoughts Tried to illustrate potential for using rich internet data to study market mechanisms, consumer behavior and competition One challenge is to find useful and “scalable” empirical strategies Simply applying machine learning techniques will often not be enough to answer typical economic research questions Illustrated just a few examples of strategies (“experiments” and “tax surprises”), and several applications: consumer demand, price dispersion, tax sensitivity. Much more can be done. Final note: “scalable” / “big data” strategies will almost surely not be as clean as a carefully designed field experiments or quasi experiments, but the benefits are clear One advantage of “big data” is that the importance of many plausible concerns could be empirically evaluated
  29. Backup slides

  30. Do Consumers Treat All “Prices” the Same? Internet markets are a great place to test behavioral hypotheses about consumers: an important one relates to price “salience” Do consumers internalize shipping fees? Hossain-Morgan (2006), Brown-Hossain-Morgan (2009), Tyan (2005), provide evidence that shipping fees are not fully internalized Shipping fees aren’t displayed in default search results, although “free shipping” is noted Consider experiments that have multiple auction listings that have different (flat rate) shipping fees Report estimates based on 7,000 experiments, around 120,000 auctions Fees have minimal effect on sale probability, so focus on seller revenue
  31. Effect of Shipping Fees Effect of incremental shipping fee Effect of “Free” Shipping Increase of $1 in shipping fee leads to $0.18 increase in price + shipping in baseline sample => shipping fees are only 82% internalized.
  32. Cross-state substitution: results Implication: state tax increase of 1% shifts 5% of in-state purchases to out-of-states (holding total online purchases fixed).
  33. Online-offline substitution: panel Implication: state tax increase of 1% increases online purchasing by state residents by around 1.8%. Conversely, taxing online purchases at state tax rate of x% reduces online purchasing by 1.8*x%.
More Related