ECONOMIC CARTOGRAPHY: MAPPING THE PRODUCE MANIFOLD
Economic cartography is a term that I want to hear discussed more often. We live in a world where instantaneous snapshots of the global consciousness are accessible with only a click. But why is it so hard to track the price of butter with similar fidelity?
To be clear, I don’t just mean choropleths mapping the relative price of Big Macs across the globe; but rather, bringing the full power of creative data visualization to bear on producing high-rez, realtime visualizations highlighting changes in demand, availability, and seasonal correlations.
Premise is building a global, human-directed machine capture infrastructure for economic data. We are focused on delivering real time monitoring and analytics to a fast-growing customer set in commercial, financial and government sectors. One critical benefit of our pipeline is the ability to collect, process and analyze ‘analog’ information—effectively, we are indexing the real world and making it available for econometric analysis, at scales, resolutions and latencies previously unattainable.
Read more at premise
‘HACKING’ CHINESE INFLATION IN REAL-TIME
The team at Premise has built an alternate method for tracking Chinese “inflation”: pointing infrastructure at hundreds of thousands of Chinese consumer prices across web and real world shelves. The advantage to our approach is the ability to spot emerging trends in real or near-time. Typical headline data releases happen monthly at 6-8 week lags for country-level “inflation” data. Countries such as the US, Brazil, China all take a six or seven decades’ old manual survey methodology. The problem here is that technology has immeasurably sped up economic life over the past decade, while our institutions continue to measure according to an old regime.
Read more at premise.
economics 2.0: data science meets the dismal science
Online activity at any moment in time [is] a snapshot of the collective consciousness, reflecting the instantaneous interests, concerns, and intentions of the global population [Duncan Watts / ”predicting the present”]
As more real-world social and consumer activity begins to move online, a natural question to ask is what kinds of economically relevant information can be mined from the growing mass of news articles, blog entries, product and venue reviews, prices, status messages, Web searches, location-based venue checkins, page views, ad impressions, etc., that are becoming available? That is, can we translate quantitative measurements about the current state of the Web into a meaningful measurements about the state of the world?
A small but growing body of literature exists on this subject: e.g., high Yelp ratings lead to higher restaurant revenue, Twitter political sentiment predicts polling data, aggregate measures of blog anxiety predict S&P returns, etc (see below for a more exhaustive list).
Regardless of whether hedge funds are using Twitter to predict the future or not, the site itself certainly functions as a fast, responsive, fine-grained gauge of the current human condition. So a hypothetical Twitter consumer sentiment index (say, tweets containing positive sentiment words along with “bought” or “purchased”) or Foursquare unemployment index (measured, say, by coffee-shop checkins during working hours) is certainly plausible. And indeed darker data pools like the user intent captured by search queries have potentially even more economic relevance.
Traditional macroeconomic measurements, however, are slow and coarse-grained, little pinpricks of insight at monthly intervals, and require significant human effort to scale. In contrast, collecting realtime sentiment data from Twitter or scraping prices from millions of shopping websites requires far less human capital, but incurs a significantly higher data processing and storage cost.
Enter data science.
Historically data science on the Web has contributed large-scale predictive models in two broad categories: (1) monetization (read: advertising optimization) and (2) content recommendation / analysis. These are both first-order problems in service of building and scaling Web applications. As the complexity and interactivity of such applications increases, the higher-order problem of turning data exhaust into economic insight becomes more important, and many of the same tools are relevant.
Unsurprisingly, a lot of smart people have been thinking about this, and what follows is a small primer. In addition to the large institutional players like Reuters and Bloomberg, companies like recordedfuture and palantir are actively engaged in predicting the future with this data.
The following list contains mainly relevant academic work, including work on mining old-media, since the tools are fundamentally similar. If you know of any citations I’m missing please let me know.
- Sentiment cues in film reviews predict box office success, moreso than approaches based solely on movie metadata [Joshi et al.; 2010]; also [Zhang and Skiena; 2009]
- The volume of movie-related posts on Twitter predicts opening weekend revenue (moreso than sentiment cues) [Asur and Huberman; 2010]
- Yahoo! search query volume predicts: (1) opening weekend box office ticket sales, (2) first month revenue for video games, and (3) the weekly popularity of top songs. The effect is stronger for box office tickets than for music, indicating that search behavior in the former case is more oriented towards actually purchasing movie tickets, rather than finding ancillary information such as lyrics [Goel et al.; PNAS 2010]
- Mentions in blog posts correlate with movie revenue [Mishne and Glance; 2006]
- Increases in blog mentions forecast spikes in the Amazon sales rank of books [Gruhl et al.; KDD 2005]
- A one-star increase in Yelp rating leads to a 5-9% increase in revenue for independently-owned (non-chain) restaurants. Chain restaurants have declined in market share as Yelp popularity has increased; i.e. Yelp is replacing traditional forms of reputation brokering [Luca; HBS 2011]
Political Sentiment and Consumer Confidence
- Twitter sentiment cues correlate strongly with surveys of US public political sentiment and consumer confidence, and are potentially predictive up to 30-days forward. Furthermore this correlation appears to be increasing over time as Twitter gains more traction [O’Connor et al.; ICWSM 2010]
- Relative political party mention volume on Twitter corresponds closely with polling data in German elections [Tumasjan et al.; ICWSM 2010]
- Search query volume for flu-related symptoms can forecast US regional flu outbreaks [Ginsberg et al.; Nature 2008]
- Influenza-related Twitter message volume correlates highly with US national flu statistics [Culotta; 2010]
- Aggregate anxiety, worry and fear on LiveJournal is Granger-causal of S&P 500 returns. However, this effect is no not significant when additionally regressing against VIX [Gilbert and Karahalios; ICWSM 2010]
- Mentions on CNBC trigger significant increases in company stock price regardless of the sentiment of the mention. [Shabani; 2011]
- Negative media sentiment can forecast short term movements in stock price. Both strong positive and negative sentiment predict high trading volume. These effects are transient, followed by a reversion to fundamentals. [Tetlock; Finance 2007]
- Abnormally high message activity on ragingbull.com (1999-2000) correlates with abnormal industry-adjusted returns and high trading volume; however this activity is not predictive. [Tumarkin and Whitelaw; 2000]
- Likewise message board activity on engadget.com is correlated with market activity, but not predictive. [De Choudhury et al.; 2008]
- Google maintains a domestic trends index tracking economically relevant aggregate search traffic across a diverse set of sectors, ranging from auto and home sales to US unemployment claims [Choi and Varian; 2009]
- The Billion Prices Project maintains a bottom-up, realtime consumer price index based on automatically extracting and classifying prices from online retailers [Cavallo and Rigobon; 2011]
Why generic machine learning fails
I get pitched regularly by startups doing “generic machine learning” which is, in all honesty, a pretty ridiculous idea. Machine learning is not undifferentiated heavy lifting, it’s not commoditizable like EC2, and closer to design than coding. The Netflix prize is a good example: the last 10% reduction in RMSE wasn’t due to more powerful generic algorithms, but rather due to some very clever thinking about the structure of the problem; observations like “people who rate a whole slew of movies at one time tend to be rating movies they saw a long time ago” from BellKor. Flexible classification frameworks only helped in as much as they were capable of handling additional features. This is in direct contrast to companies like kaggle, by the way, which are distilling and bottling machine learning knowledge and talent at a fantastic rate.