AgPa #7: Spotify Streaming and Stock Returns

Music sentiment and stock returns around the world (2021)
Alex Edmans, Adrian Fernandez-Perez, Alexandre Garel, Ivan Indriawan
Journal of Financial Economics, In Press, Corrected Proof, URL

Building on last week’s introduction of alternative data, this week’s AGNOSTIC Paper examines the role of music sentiment in the stock market. What sounds like statistical hocus-pocus is part of an important question. Do other factors than rational information drive stock markets?[1]One of the authors, Alex Edmans, elaborates on this question in an interview with HarvardBusinessReview. This is the old debate of market efficiency, the idea that the current price of a security reflects all available information about it.

The authors (unsurprisingly) do not agree with this notion. They show that stock returns tend to be positive when people listen to happier music on Spotify. Hence, public sentiment seems to have substantial impact on returns. I like the paper for its creative use of alternative data and its clean methodology. But to be honest, I was also somewhat skeptical when I first heard about it. If you analyze enough data, you will eventually find some significant correlations with stock returns. This problem is called data-mining or data-snooping. However, the authors present an intuitive economic rationale and rigorously test their hypotheses in various robustness checks. Therefore, I believe that they indeed found an important mechanism that is active in stock markets. As usual, I divide this post into the following parts and you can use the list to navigate.

Everything that follows is only my summary of the original paper. So unless indicated otherwise, all tables and charts belong to the authors of the paper and I am just quoting them. The authors deserve full credit for creating this material, so please always cite the original source.

Setup and Idea

Market efficiency is a fierce debate among economists and practitioners. I will not comment on all nuances of the discussion but one important aspect is sentiment. Baker and Wurgler (2007, Abstract), two pioneers in this field, define sentiment as “belief […] that is not justified by the facts at hand.” For example, after a bad night of sleep I am more pessimistic although the world around me hasn’t changed. That’s sentiment and since it affects my actions[2]In a rational economist’s world it shouldn’t affect my actions and I work hard to avoid it. But we are all just humans…, it has some real-world effects.

The primary challenge of all research on investor sentiment is to come up with empirical proxies. How should we capture the mood of investors appropriately? As usual, this is not easy because there are a lot of things at work simultaneously. That is obviously the point where the paper kicks in. As we will see, music sentiment is a measure that works quite well and has some distinct advantages.

Such effects obviously doesn’t fit into the world of rational efficient marketers. However, it is important to understand that they are not questioning the existence of sentiment by itself. They just say that if sentiment pushes prices away from their fair value, some rational investors will exploit this quickly. The sentiment-driven investor loses, the fastest arbitrageur wins, and nothing changes for the average (slow) investors. According to this logic, the market still reflects all information despite distortions from sentiment. The question about market efficiency is therefore not if mispricings exist, but how long they exist and if the average investors can exploit them.

Data and Methodology

The key innovation of the paper is novel data on music sentiment of Spotify streams. Since 2017, Spotify provides daily statistics for the 200 most-streamed songs in each country. A stream is counted only if the user listens to the song for more than 30 seconds. Songs that are immediately skipped by the user are therefore excluded.

Spotify also provides a sentiment score called Valence. Valence measures the sentiment of each song and ranges from 0 (negative) to 1 (positive). The score is based on a proprietary machine learning model originally developed by The Echo Nest.[3]Some more information about it here. The Echo Nest was later acquired by Spotify. It is fairly sophisticated and not only includes the sentiment of lyrics, but also other features like melody or beats per minute. The details are proprietary, so the analysis is not fully transparent at this point and we have to rely on the data scientists from Spotify.

The authors use this data to calculate a stream-weighted average valence (SWAV) for each country and day. Although it sounds complicated, it just means that the most-streamed songs receive higher weights in daily averages. I suspect that streams are heavily concentrated to a few top-hits. Stream-weighted averages are therefore more reasonable. The following chart shows the average SWAV between 2017 and 2020 in 40 countries.

Figure 1 of Edmans et al. (2021).

The higher (lower) the score, the more positive (negative) the music in a particular country. In line with feel-good Latin beats, Spaniards and South Americans clearly listen to more positive music than the rest of the world (SWAV >0.6). In contrast, most Asian countries, the US, and Canada seem to favor more negative songs (SWAV around 0.45). Europe is somewhere in between, but it also depends on the country.

The authors use this music-positivity as a proxy for public sentiment within a certain country. Compared to other measures, this data has several advantages. First, it is available at a high frequency (daily) for many countries. Some important ones are missing (especially China), but 40 countries is still a lot. Second, music-sentiment does not depend on the language which makes global comparison much easier. Finally, it is continuously available and does not depend on specific events.[4]For example, the authors cite other studies that measure sentiment after events like terrorist attacks or the elimination of a country in the soccer world cup.

To further validate their measure, the authors compare it to other proxies for public sentiment. For example, they show that people listen to happier music during positive months. Researchers found that on average, people are happier during January and March because of the new-year-spirit and the beginning of spring.[5]The “on average” part is important. Personally, I hate the beginning of spring because I am allergic to pollen. But January and March are only two examples. There is much more research on positive and negative months. They also compare music sentiment to the weather. Again, prior research identified that people tend to be less happy on cloudy days (You don’t really need science for that, right?).[6]Yes, there is data on daily cloud average across countries. Living in the big-data-age has some advantages. Finally, they look at music sentiment during COVID lockdowns. In countries with stricter lockdowns, people seemed to be less happy and listened to more negative music. Overall, music-sentiment appears to be a very robust proxy for public sentiment.

To examine the relation to stock returns, the authors take weekly changes of the SWAV scores in each country. They use changes because the absolute level differs across countries and time. Furthermore, sentiment-shifts should actually be more important for stock returns than the sentiment-level. Going from optimism to pessimism probably triggers more negative returns than remaining pessimistic. Although the authors always use the weekly change for their analyses, I will just refer to music sentiment for brevity.

The authors relate music sentiment to daily returns of MSCI country indices in USD. This is just one time series per country, so there shouldn’t be any data issues. The only remaining problem may be look-ahead bias. Spotify releases the top songs daily, but the list may not be available during market-opening-hours of the same day. The authors solve this issue by using weekly returns and music sentiment.

Important Results and Takeaways

The following table summarizes the main result of the paper. Music sentiment is positively related to weekly stock returns. When people listen to more positive music, global stock returns tend to be higher. This effect remains statistically significant after controlling for various other variables like the return on the MSCI World, the VIX index, weather, and several proxies for economic policy.

Table 3 of Edmans et al. (2021).

Notably, the effect reverses when we look at the predictive power of music sentiment. In Panel B, the authors regress weekly stock returns on the music sentiment of the week before and find a negative relation. Again, this effect remains statistically significant after all control variables.

This is consistent with a mispricing-reversal-pattern. Consider for example a positive change in music sentiment (people listen to happier music). The behavioral story goes like this. First, sentiment-driven investors produce positive returns by bidding up the price. Subsequently, some smart investors recognize that this was just sentiment and correct the price movement by selling.

It is completely okay to be skeptical about this story and the results. Of course, not all investors listen to music. But this is not a weakness of the paper. Although music sentiment doesn’t capture the mood of all investors, it is a good proxy for public sentiment. And the analysis strongly suggests that the latter is important for markets.

But to address the skeptics, the authors provide a lot of robustness checks. First, they control for seasonal effects by including calendar-month dummies. Second, they exclude the top 50 songs to ensure that music-sentiment is not driven by a few top-hits. Third, they re-run the analysis another 40 times and exclude one country at a time. This ensures that the effect is not driven by one particular country. And finally, they also examine the relation on the daily level. In all those robustness checks, the impact of music sentiment remains highly significant.

Of course, you can never proof a hypothesis by adding up robustness checks. But I think we get enough evidence to believe that the effect of sentiment is real and not due to data-mining.

Music sentiment is more important in less efficient markets

In the next part, the authors examine how “limits to arbitrage” affect the impact of music sentiment. The idea is simple. Suppose a stock is trading at $9.90 but is actually worth $10.[7]Let’s just assume we know that the true value is $10. In practice, this is a bit harder if not impossible. There are only two reasons for the deviation. First, nobody recognizes the opportunity and you have found a way to make a profit of 10 cents per share. Congratulations! Alternatively, there are some “limits to arbitrage” that prevent smart investors from stepping in. For example, the stock could be illiquid such that transaction costs exceed the profit. In such a situation, the price remains at the “incorrect” $9.90 and the market is not fully efficient.

The authors test the effect of music sentiment in markets with such “limits to arbitrage”. To do that, they exploit the unique setting of the COVID-crash. In March 2020, some countries imposed severe trading restrictions to stabilize markets.[8]We can debate if this is a good or bad idea, but I guess you know my stance on that… For example, France banned short-sales for two months and Australia even imposed an upper limit of trades per day. We would expect that sentiment-mispricing is stronger in such markets because it is harder for arbitrageurs to correct it.

The authors indeed find this pattern. The effect of music sentiment on stock returns is significantly higher in markets with more trading restrictions. However, I would like to mention that the outbreak of COVID in early 2020 was a very special situation. Therefore, I think it is somewhat hard to generalize from that period. On the other hand, it is fair to assume that short-sale-bans will never help to correct mispricings.

To further stress-test their analysis, the authors apply music sentiment to other settings. Statistically, that’s an out-of-sample test and our best weapon against data-mining.

First, they look at mutual fund flows and obtain data on 8,392 equity funds in 31 countries from Morningstar. Positive net fund flows indicate that more money was invested in funds than withdrawn. This is a sign of optimism among investors. The authors find a significant positive relation among music sentiment and fund flows. When people listen to happier music, they tend to invest more money in the stock market. This is again in line with the authors’ sentiment-hypothesis.

Finally, the authors relate music sentiment to returns on government bond indices from Refinitiv. The idea is straight forward. When sentiment turns negative, investors often move from risky- to safe assets (“flight to safety” or “safe havens”). Therefore, (safe) government bond returns should be negatively related to music sentiment. Once again, the authors find this pattern in their data. However, there is no reversal-effect over the following week as for stock returns.

Conclusions and Further Ideas

In my opinion, the idea of this paper is just brilliant. The authors do world-class research (the Journal of Financial Economics is among the Top-3 globally) but the idea and main result remains simple and easy to understand. I also like the methodology and structure of the analysis. Based on economic intuition, the authors first form a hypothesis (sentiment affects stock returns). Then, they construct an empirical proxy to test it (music-positivity is a robust proxy for sentiment). And finally, they try a lot of things to kill their hypothesis (robustness checks and out-of-sample tests). That’s the scientific method in action.

What can we do with the results in practice? An obvious application is a sentiment-driven timing model for stock- and bond-indices. It would be interesting to see the backtest of a trading strategy that exploits the documented mispricing-reversal effect. Unfortunately, the authors haven’t done this in the paper. Another application is now-casting of sentiment. Investors always try to guess the current mood of the market and look at a lot of signals. With it’s high frequency and global coverage, music sentiment is certainly a useful addition for that.

All of this sounds promising, but I think academia is somewhat behind the curve here. The large quant-firms traded such short-term patterns for decades and probably have way more sophisticated measures for sentiment. This is no critique of the paper. Quite on the contrary, I believe it’s great how the authors isolate the underlying effect. But in the end, the question about market efficiency remains unchanged. Even if music sentiment predicts future stock returns, can the average investor really profit from it? In most cases, the answer is probably no. So even with irrational sentiment, financial markets remain heavily competitive and it is not easy to earn abnormal returns.

This content is for educational and informational purposes only and no substitute for professional or financial advice. The use of any information on this website is solely on your own risk and I do not take responsibility or liability for any damages that may occur. The views expressed on this website are solely my own and do not necessarily reflect the views of any organisation I am associated with. Income- or benefit-generating links are marked with a star (*). Please also read the Disclaimer.