The Conflicting Links between Forecast-Confidence and Trading Propensity

Abstract

While finance studies suggest that forecast-confidence motivates trading, the experimental findings regarding the confidence-trading links are inconclusive and statistically weak. Attempting to bridge the gap, we modify the standard interval forecasting task to measure forecast-confidence more directly. The adapted task is utilized to test the confidence-trading correlations at the attitudinal level and in specific scenarios. The attitudinal test surprisingly reveals that forecast-confidence negatively correlates with the inclination to churn one’s stock portfolio, although confidence in profitability indeed boosts the willingness to trade particular stocks. The attitudinal correlation is endogenous, brought by opposite personality and competence effects on confidence and trading.

Keywords:

• Forecast-confidence
• Trading propensity
• Personality
• Overprecision

JEL CLASSIFICATIONS:

• D8
• C9
• G1

Acknowledgement

We thank Ido Erev, the Technion R&D foundation, and the research authority at COMAS (College of Management Academic Studies), for supporting this research. The study was presented at the 2017 SPUDM conference at the Technion, the IAREP 2017 meetings in Israel, the 2017 European ESA meetings in Vienna and seminars at the Gothenburg Research Institute, the Technion, and Ben-Gurion University. We have benefited from exchanges with Klajdi Bregu, David Budescu, Ido Erev, Yoav Ganzach, Sven Hemlin, Magnus Jansson, Ido Kallir, Doron Kliger, Ferdinand Langnickel, Christoph Merkle, Orkan Nadirov, Nigel Harvey, Dilek Onkal and Eldad Yechiam.

Notes

1 Confidence and overconfidence positively correlate (see, for instance, the classic calibration curves in Lichtenstein, Fischhoff, and Phillips 1982) but still represent distinct constructs. Moore and Healy (2008) separate between overestimation of individual abilities, overplacement relatively to others, and overprecision in subjective beliefs. Several papers show that trading propensity may increase with overplacement or overestimation (Glaser and Weber 2007; Michailova and Schmidt 2016; Bregu 2020), but we presently focus on forecast-certainty and overprecision. The terms forecast-confidence and forecast-certainty are used interchangeably.

2 Nosic and Weber (2010; discussed later) is an exception.

3 In Biais et al. (2005) miscalibration links with lower earnings in experimental markets. The current study is not designed to test the forecast-overprecision effect on earnings from trade (see the concluding Discussion).

4 We adopt the Merkle (2017) definition of volatility-based overprecision. Other studies use slightly different definitions. Nosic and Weber (NW; 2010), for example, use the negative (-perceived volatility)/(empirical volatility) ratio. When the prediction target is fixed across the sample (as in NW), the empirical volatility estimates are fixed, so differences in overprecision only follow from differences in perceived volatilities.

5 Interval-evaluation tasks were applied in few judgmental psychology studies (Murphy and Winkler 1974; Winman, Hansson, and Juslin 2004; Teigen and Jørgensen 2005; Speirs-Bridge et al. 2010), but we are not aware of financial forecasting studies employing a similar method. More detailed discussion of FAAT and the background judgmental psychology literature is provided in Sonsino, Lahav, and Levkowitz (2020).

6 A loss function for the $\alpha$ quantile of X is $|x-q|·(\alpha ·1_{\left\{x>q\right\}}+(1-\alpha )·1_{\left\{x\le q\right\}}),$ where $q$ is the elicited $\alpha$ quantile and x is the realized X, while a quadratic scoring rule for the likelihood p of event E is $p^2·1_E+ (1-{p)}^2·1_{E^C}$ (Gneiting and Raftery 2007). The quadratic score takes only 2 values, while the loss function for $\alpha$ varies with $x.$

7 Similar results emerge when the sample is extended to N = 93 subjects with 1–2 errors or omissions and the problematic tasks are ignored. The results for the extended sample are presented in Web supplement 6.

8 The probability of winning the 100 NIS was 100% for prediction errors (|F-r|)) smaller than 1%, 98% for errors smaller than 2%, etc. The quadratic scoring rule 100*[1-(1-P)²], where P represents the likelihood assigned to the realized event (hit or miss) in decimal form, was used to derive the winning probability from the likelihood assessments. When drawing the payment assignment, we also drew a 1–100 integer representing the winning threshold for the class. The subjects received the 100 NIS when their (payment assignment) winning probability exceeded the threshold.

9 To decrease the likelihood of zero purchase or sell, the BUY (SELL) assignments dealt with stocks that showed positive (negative) returns at the first half of 2016. The larger (smaller) risk-free rate in the BUY (SELL) problems similarly aimed at decreasing the chances that subjects would choose to buy/sell the 100,000 NIS.

10 Throughout the paper, we use the sign-test for one sample hypotheses and the Pitman test for between samples comparisons. We report Spearman correlations, using randomizations to test significance. Significance levels are 1-tailed. We use “average” for within-subject statistics and “mean” for the cross-sample statistics, except for using terms such as “higher than average” in the usual sense.

11 More details are provided in Sonsino, Lahav, and Levkowitz (2020; study 1).

12 Our verbal risk-receptiveness measure is similar to the SOEP measure explored by Dohmen et al. (2011); Crosetto and Filippin (2016). In addition, Q2 presented a multiple price list task (Andersen et al. 2006) where subjects select between a fixed lottery paying 100 or 200 with equal 50% probabilities and certain payoffs that increase from 100 to 200 in increments of 10. The correlation between the verbal RR and the price list switch point is 0.44 (p < 0.01), but the switch point shows smaller predictive power for TP (correlation 0.3) and loses significance when RR is controlled.

13 Positive correlation between openness and risk receptiveness is reported in few preceding studies (e.g., Becker et al. 2012). The evidence on personality effects on financial decision, however, appears inconclusive (see Furnham 2020 survey) and we could not generally connect the current results to preceding research. Regarding openness, note that Schaefer et al. (2004) find positive correlation between openness and confidence in binary choice general knowledge problems, while we find negative correlation between openness and forecast-confidence. Note also that our Neuroticism correlations with TP and CONF are of the same sign as the E, C, O correlations, although principal component analysis confirmed the existence of a big one trait that negatively loads on N (Musek 2007).

14 In general, the evidence regarding competence effects on overprecision is mixed. In Glaser, Langer, and Weber (2013), for example, professionals exhibit stronger overprecision than students, while Gloede and Menkhoff (2014) find that overprecision decreases with investment experience.

15 The split could build on alternative indices; e.g., similar results emerge when the subjects that score high in C, O and N (normalized scores larger than –0.75 in all three traits; N = 37) are separated from others.

16 Similar results emerge assuming that the return is uniformly distributed around F. Confidence in profitability could be elicited directly by adding a fourth step to the FAAT, but this could produce spurious correlations between confidence in profitability and the buy/sell amounts.

17 The J-test may bring 4 outcomes: The F model is rejected for the CONF + model, the CONF + model is rejected for the F model, neither model is rejected or both models are rejected. See, for example, Bellemare and Barrett (2006).

18 The total BUY/SELL amounts of the two groups do not differ significantly (means 46K vs. 44K; p = 0.35) and their average CONF + are similar (means 59 vs. 63; p = 0.14). The mean TPs are 43 and 52 (p = 0.02). The significance of the differences in responsiveness is weaker (0.16 vs. 0.52; p = 0.12) when the subjects with median MIN_NH CONF = 70 are included.

19 Predictability is generally stronger for longer horizon and less volatile return targets so that FAAT can generate meaningful forecast-overconfidence measures in other applications.

20 The regression results are slightly weaker when familiarity is replaced with the competence index. We only disclose T-values since the marginal effects are not informative in ordinal Probit regressions.