15 June 2020
The efficient market hypothesis (EMH) and financial economic theory assume all market participants are completely rational. The EMH argues that stock prices reflect all publicly available information related to a stock’s future cash flows and risk, and that when new information is released to the market, the stock price will change to reflect this new information. If the EMH always held true, there would be no opportunity in the market for outperformance, as there would be no way to obtain any long-term edge. Enter behavioural biases.
In reality, market participants are humans and therefore do not always act rationally. Asset pricing theories struggle with human behaviour, which is often prone to biases. It is these biases that drive the divergence between theory and reality.
The rise of behavioural finance and market participants’ increasing acceptance of irrationality by market participants has given rise to the concept of “sentiment” in investing. There are typically two schools of thought as to why sentiment anomalies are observed: risk factors and investor biases.
In the Macquarie Systematic Investments team, we subscribe to the theory that investor biases dominate, and believe that we can systematically exploit these investor biases to generate outperformance.
Consider the following two questions:
- Is this car worth $US20,000?
- How much is this car worth?
Your guess on the second question was affected by the suggestion of the price for the car, which led you to consider the $US20,000 price, and then insufficiently adjust from that starting point. In fact, the 2020 Kia Rio shown here has a manufacturer-suggested retail base price of around $US16,675, according to motortrend.com.
This is an example of the human bias known as “anchoring.” Anchoring is when an investor relies too heavily on a prior estimate of value, or when one considers a value for a quantity before estimating that quantity. Anchoring and herding (discussed later) are key investor biases underpinning the ability to invest to capture sentiment.
Below we further discuss sentiment in terms of its measurement and how our investment process looks to exploit it. Sentiment is one of the three key pillars of the Macquarie Systematic Investments team’s investment process – the other two are value and quality.
What is sentiment?
Sentiment is the tendency for asset prices to continue moving in the same direction as before (that is, a company that has outperformed recently will continue to outperform, and vice versa). Often, the asset price deviates from what may be expected based on rational economic pricing.
Unlike other market metrics (such as valuation ratios), one challenge of utilising sentiment is that there are many possible ways to indirectly measure it. These indirect measures are used to represent some manifestation of the sentiment effect but are never able to explicitly measure the true market sentiment. Below we look at two of the most long-standing measures of sentiment: price momentum and fundamental momentum. We then look at a more recent approach that extends the analysis by decomposing share price performance into its principal components.
Price momentum is defined as a stock’s price return over some period of time. Jegadeesh and Titman (1993) highlighted the anomaly of positive excess returns to buying past winners and selling past losers. They hypothesised that the anomaly is a product of investor underreaction (via anchoring) to new information, and a slow incorporation of relevant information into asset prices.
While price momentum strategies have been a cornerstone of typical systematic investing processes due to their continued efficacy, the strategy can be exposed to rapid drawdowns during times of high volatility and market uncertainty.
Figure 1 shows the long-term spread of mean returns between stocks with high momentum scores and stocks with low momentum scores. Whilst over the long-term we can see that the spread has remained positive – implying a positive return to companies with high momentum – during market dislocations such as the 2008-2009 global financial crisis, there was a period of extreme behaviour in which the mean return spread suffered strong inversion.
Figure 1. Spread in median return between top and bottom quintiles from a portfolio sorted on 6-month price momentum
Based on the S&P/ASX 300 Index
Sources: Macquarie Investment Management and S&P.
Figure 2 shows that the anomaly during the global financial crisis was driven by mean reversion in stocks with poor momentum scores – prior losers had a period of strong outperformance and mean reversion.
Figure 2. Equally weighted quintile returns from a portfolio sorted on 6-month price momentum
Quintile 1 = bottom scoring stocks, based on the S&P/ASX 300 Index
Sources: Macquarie Investment Management and S&P.
Such behaviour illustrates that sentiment effects can become dislocated during periods of uncertainty. The period of uncertainty is then typically followed by a period of recalibration. It is during these periods that the sentiment factor is less effective, as investor behavioural biases shift away from chasing winners and into a more defensive approach to stock picking.
Fundamental momentum or earnings momentum
Earnings are typically reported on either a semiannual or quarterly cycle, with companies also delivering ongoing updates on their expectations for what reported earnings will be. A key component of this earnings reporting cycle is the influence of broker forecasts on a stock’s price, and the general sentiment behind these forecasts.
Manifestations of systematic biases in earnings include the over/under reaction to revisions in earnings forecasts and companies failing to meet or exceed their own forecasted earnings as well as market expectations. This effect can be particularly observed in broker forecasts, where there is strong herding behaviour. There is career risk in an analyst producing an earnings forecast that deviates substantially from consensus. Thus, what tends to be observed is a statistical “crowding” effect in earnings forecasts, followed by incremental uplift in the forecasts as certainty increases.
All this can lead to a similar momentum effect, in that stocks with previously strong earnings are likely to exhibit strong earnings into the future. Examples of measures of earnings momentum include:
- changes in reported earnings per share (EPS)
- changes in broker-forecasted EPS (consensus)
- number of upward and downward revisions.
A further evolution: Dissecting sentiment
Taking the more traditional concepts of systematic momentum and fundamental momentum above, we can then extend our analysis to decompose the sentiment of a stock into its principle components. It is possible to decompose total stock return into its price and dividend component:
Total stock return = price return+dividend return = ΔEPS+ΔPE multiple+interaction+dividend return
This decomposition provides insight into what component of a stock’s return is driving the momentum anomaly. For example, there could be two stocks with an equivalent return over a six-month period, but one stock’s return has been driven by upward revisions in earnings while the other’s return has been driven by price-to-earnings (P/E) multiple expansion (that is, a change in investor appetite for a company or its characteristics). From a sentiment perspective, which stock should we prefer?
Our team’s research shows there is greater persistence in excess returns for companies with upward revisions in earnings. Companies driven by P/E expansion tend to revert over time. As a result, the team’s process prefers companies with strong fundamental momentum – those for which share-price performance has been supported with upward revisions in earnings.
Keeping sentiment relevant
With the exponential growth of data, computing power and the resurgence of machine learning, new opportunities for equity managers to capture, understand, and exploit sentiment have emerged in recent years.
Below we provide a high-level overview of some of the more recent additions to our process.
Applications of machine learning
Machine learning momentum
One of the key anecdotal observations of the momentum anomaly is that it works, until it doesn’t. Momentum relies on some underlying trend, which forms part of a stock’s price and persists over time. It is well known that there is also mean-reversionary behaviour in stocks (source: Barberis, Shleifer, and Vishny, 1998). Thus, we can think of stocks as belonging to two regimes: trending and reverting. We can seek to find a way of identifying whether a stock is currently operating in one of these regimes. There are many potential approaches to solving this problem, largely stemming from the field of pattern recognition and machine learning.
Deep learning to forecast earnings
One of the key requirements for applying earnings momentum to a universe of stocks is broker coverage. Brokers provide forecasts for a variety of key metrics, and these metrics can then be used as part of the investment process to understand a stock’s future earnings potential. But what happens when a stock isn’t covered by any broker? This requires an analyst to build a financial model to forecast earnings. While doable, this is not scalable across large universes of stocks. Enter deep learning.
Alberg and Lipton (2018) demonstrated that by using previous period company fundamentals, it was possible to use neural networks to forecast company fundamentals in the future. The application of this sort of analysis is quite substantial, since with just a home desktop computer, it is possible to build forecasts for entire universes of stocks from publicly available information. Naturally, at the most basic level, nuances associated with each company and sector will be missed, but the scalability of this technology means it is possible to build sector- or even company-specific models that subsequently run automatically when new data is available.
There has been considerable attention given to applying machine learning techniques to text-based data such as news articles and earnings-call transcripts. One of the key areas of research has been attempting to extract the sentiment of a body of text and link this with the current sentiment of a stock. This sort of alternative way of thinking about sentiment and using data will likely continue to grow, as it provides a methodology for linking qualitative information surrounding stock markets into a more concrete quantitative measure, such as sentiment.
Sentiment is a key concept that has risen out of a shift away from traditional financial theory into understanding how human behavioural biases are incorporated into the pricing of assets. As highlighted in this paper, humans are far from purely rational agents seeking to optimise a utility function. Macquarie Systematic Investments believes having a strong understanding of these biases is a key to systematic investing. As academia and industry continue to evolve their understanding of behavioural finance, and as access to new data sources and computational techniques continue to grow, we believe evolution of the sentiment factor will continue to give us an edge.
Alberg, J., and Lipton, Z. (2018). Improving Factor-Based Quantitative Investing by Forecasting Company Fundamentals. arXiv. Retrieved from https://arxiv.org/abs/1711.04837
Barberis, N., Shleifer, A., and Vishny, R. (1998). A Model of Investor Sentiment. Journal of Financial Economics, 49, 307-343.
Daniel, K., Hirshleifer, D., and Subrahmanyam, A. (1998). Investor Psychology and Security Market Under- and Overreactions. The Journal of Finance, 53(6), 1839-1885.
Hong, D., Lee, C. M., and Swaminathan, B. (2013). Earnings Momentum in International Markets. Research Collection Lee Kong Chian School of Business.
Jegadeesh, N., and Titman, S. (1993). Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency. The Journal of Finance, 48(1), 65-91.
Tversky, A., and Kahneman, D. (1974). Judgment under Uncertainty: Heuristics and Biases. Science, 185(4157), 1124-1131.