Anticipating the next NFP Surprise? Mind the Sentiment!

Last week, the US economy was expected to have added nearly 1m jobs in April, but what came out on the 7th of May was a meager +266k. The surprise was all the more shocking, given that the standard preview checklist was largely positive with only the employment sub-index of the ISM Manufacturing PMI clearly pointing to a potentially weaker than expected print (cf. Figure 1). Could news sentiment have helped us see the negative NFP surprise coming and is it something we should be adding to our NFP checklist going forward? To answer these questions, we use news sentiment data derived from RavenPack Analytics.

FIGURE 1 - US Jobs Report Pre-release checklist - May 7th, 2021 - Source: FXStreet

RavenPack has developed proprietary natural language processing (NLP) algorithms to identify in real-time almost 7000 events in thousands of news articles daily. We leverage this granular event taxonomy to filter news related to the Labor Market (employment, unemployment, jobless claims, nonfarm payrolls, hirings, layoffs, etc. events).

To create our Labor Market Sentiment indicator, we use the RavenPack Composite Sentiment Score (CSS). CSS captures the sentiment of a given document constructed by combining various natural language processing techniques. The direction of the score is determined by looking at emotionally charged words/phrases and has been trained to match their short-term impact on large-cap stocks share prices.

Daily sentiment can be noisy, we therefore take a monthly exponential moving average of the daily CSS scores so that we smooth out the noise while still giving higher weights to sentiment on most recent days. We then standardize it using a rolling window of the past 3 years to ensure that negative (positive) values can be interpreted as below (above) trend sentiment.

FIGURE 2 - Labor Market Sentiment- Source: RavenPack

As we can see from Figure 2, our Labor Market Sentiment index had been steadily falling since early April and had turned negative ahead of the release.

Could these developments have helped us predict a negative surprise?

To answer this in a systematic way, we gathered data on NFP releases and consensus estimates from August 2004 to March 2021, and backtested our Labor Market Sentiment index.

First of all, we look at whether conditioning on the level of the sentiment changes the distribution of positive vs negative NFP surprises. Figure 3 shows that it does: observing a negative (positive) sentiment on the day before the release significantly increases the probability of consensus being overly optimistic (pessimistic).

FIGURE 3 - Negative Sentiment on the day before the release increases the likelihood of a Negative Surprise, and vice versa - RED: Negative Surprise, GREEN: Positive Surprise. Source: RavenPack

Next, in Figure 4, we examine whether there are any discernible patterns in sentiment ahead of positive and negative surprises. For this, we normalize sentiment to zero on the day of the surprise and look at the evolution of sentiment beforehand.

Figure 4.a shows the sentiment paths 20 days before the positive and negative surprises using median sentiment changes, while Figure 4.b uses the average sentiment changes. Both suggest that sentiment tends to worsen ahead of negative surprises and improve ahead of positive surprises.

FIGURE 4a - Median change used for path reconstruction

FIGURE 4 - Sentiment evolution ahead of NFP surprises (Normalized to zero on Surprise Day) - Source: RavenPack

Finally, we test whether our Labor Market Sentiment indicator on the day before the release significantly changes the probabilities of observing positive and negative surprises in a regression framework.

Table 1 shows that while the consensus estimate itself and the previous NFP release do not help predict whether a positive or a negative surprise is coming up, a one-unit higher Labor Market Sentiment index makes a Positive Surprise 1.381 times more likely and a negative one 25% less likely.

TABLE 1 - Logit Regressions (odds ratios)

In sum, last week’s much weaker than expected NFP release reminded us that the standard checklists based on traditional economic indicators may not be enough when it comes to correctly assess the state of the labor market.

As discussed in this post; a news-based Labor Market Sentiment indicator helps predict NFP surprises and may therefore be a great addition to the standard toolkit