Nowcasting Signals: Risk vs Returns

Ajit Agrawal, Suchita Sridhara*, Neeraj Sudhakar

When we build our weekly revenue estimates (Nowcast) for the ~120 companies we track, clients often ask as to how to distinguish the good (vs noisy) signals. While clients are used to the concept of revenue surprise (revenue estimate minus consensus) in % terms, we use the concept of revenue surprise in terms of standard errors (standard deviation of the past errors of AKAnomics estimates), which allows us to rank our revenue estimates across companies. Which leads to the natural question as to “what standard error” should we use as a threshold above which our revenue estimate has a “high chance” of being correct. While this choice of threshold should depend upon the risk/return characteristics desired by the portfolio managers, we argue that 0.75 is a good threshold.

There are 3 factors that we believe drive the choice of this threshold:

1.       Number of active signals generated – the higher the threshold, the fewer the signals

2.       Quality of signals – the lower the threshold, the noisier the signals

3.       Market reaction – the higher the quality of the signals, the higher the potential forward returns

We address each of these factors one by one.

1.       Number of active signals generated – With a binomial distribution, a threshold of 0.75 standard deviations or more should yield a signal 45% of the time (ie around 50 signals from 120 weekly estimates on average), which is a large enough sample to build different long/short strategies. In reality, our signal distribution is not entirely normal, and has fatter tails, and this threshold yields 49% signals.

2.       Quality of signals – If our estimates followed a binomial distribution, then an estimate that is 0.75 or more standard errors away from consensus should see the outcome to be on the right side of consensus 78% of the time, and correctly predict a beat/miss 50% of the time. We see a similar pattern in our data, with our hit rates increasing with a higher threshold, and correctly predicting a beat/miss 45% of the time.

3.       Market reaction – If we observe the forward 1-week returns for our historical signals for the different threshold ranges, then we find a steady increase in the information ratio with increasing threshold, and 0.75+ yielding a decent information ratio of 2.5 (see Graph below - information ratio defined as the average forward 1-week return over the standard deviation of the returns over all signals)

With the above set of findings, we conclude that a threshold of 0.75 provides sufficient breadth of signals, of high quality, with good forward return characteristics.

*Suchita Sridhara is a Summer Intern at AKAnomics Inc, and we thank her for her contributions. Suchita is an undergrad at Wesleyan University.