The best sentiment reversal performance is obtained when net news sentiment meas-
urements are over a period of around 1.5 years. An equally weighted average over all
such events of forward 1-year excess returns relative to the S&P 500 yields 16%. More-
over, the relationship between the variable of net news sentiment measurement period
length and forward 1-year excess returns is piecewise-monotonic, thus corroborating the
presence of a genuine stock market anomaly. The Monte Carlo simulations in the final
section show that this outperformance is not market-neutral, as might be expected given
the highly sentimental impact that different economic environments have on equity
prices.
It should also be said that our study design might seem somewhat unconventional to
those steeped in the culture of financial theory. We do not use models or regressions to
‘‘explain’’ anything, nor do we control for size, book to market, or any of the standard
factors, at least in any standard way. Take the notion ofrisk, for example. Typically, one
attempts to model the risk associated with an asset by establishing correlations between
various factors, assuming that a causal relationship exists. But how appropriate is this
approach? Obviously, the post hoc ergo propter hoc fallacy is a danger in all aspects of
quantitative finance, but here it seems particularly applicable. Even if there was a causal
relationship between a factor and risk (whatever that means) in the past, it is not clear
why such a state of affairs would continue to be the case in the future. What factor could
have predicted the huge amount of risk a long-only manager would be taking in buying
financial stocks before the fabled Credit Crunch? Why not simply ask, ‘‘How much
money could I possibly lose by owning these stocks?’’ Isn’t this a more respectable
question? Our solution is to look at Monte-Carlo-style portfolio simulations and
determine the worst possible investment experience from buying a particular type of
stock.
In our professional opinion, the use of conventional investment techniques such as
factor models is not the best way to outperform the market, especially when everyone
else is abiding by the same conventions. The fact that sentiment reversals might correlate
with certain factors is mildly interesting, but without a hypothesis behind the correlation
such a research technique becomes a meaningless numbers game. We require a
convincingreasonfor a market inefficiency in order to believe that it will repeat.
9.2 The quantification of sentiment
In order to aid the study of news sentiment, DJNA has constructed an archive back to
1987 that is notsurvivor-biased. The archive is made up of tens of millions of financial
news stories, where each story is assigned numerous DJNA attribution tags, including
the historical identification of relevant companies. DJNA tracks around 27,000 his-
torical companies. It is essential in an event study to consider companies that are no
longer trading, otherwise the results have the built-in assumption that companies never
go bankrupt or get acquired. Using historical ticker symbols, we match the appropriate
CRSP security identifiers with each company-relevant news story as being the securities
that traded under those tickers when the news stories were published. This bit of
bookkeeping is a necessary preliminary to conducting a proper survivor-unbiased
study. Seeing as CRSP is the pricing source we shall use, only US companies will be
Sentiment reversals as buy signals 233