Re-examination of Crowd-sourced Earnings Forecasts from Estimize

Re-Examination of Crowd-sourced Earnings Forecasts

from Estimize

Shenghan Guo

[email protected]

Johns Hopkins University

Background──Wisdom of the crowd:

• A considerable amount of study has demonstrated thatthe estimates made by a group of people from all kinds ofbackground tend to outperform those made byprofessionals. This phenomenon is named the ”wisdom ofthe crowd”.

• There exists behavioral bias in professionals’ marketestimation such as “herding”. Fewer professionals wouldmake estimates too deviate from the majority even if theestimates are justified

• Institutional bias also exists in the sense that thefinancial institutes to which the professionals belong tomay encourage over-optimistic estimates

• Estimize is an online communityestablished at 2011, aiming at providingfinancial forecasts for key statistics, suchas EPS, revenue and etc.

• Study has shown that marketestimation from Estimize tend tooutperform estimates from Wall Streetprofessionals.

Our motivation: Examine the “wisdom of crowd”. Check if the estimatesfrom Estimize are indeed more accurate than the WallStreet estimates.Our method:Use the Wall street estimation as the benchmark andcompare the accuracy of the Estimize estimation with it.

Prior Study:

Aspects to look at:

Features of the dataset

Accuracy of the estimates

Earnings surprise

Deviation from the benchmark

Part I. Features of the

dataset

Seasonality in Estimize data:

Significant seasonality shows in the number of tickerscovered by Estimize estimation.

Compared to the version in the prior study, theseasonality appears to be more significant and theoscillation is larger.

Number of estimates as a function of the days before report:

For Estimize, the number of estimates increases asthe days before report release decreases, meaningthat there are significantly more trading signals as thetime approaching the date of report release.

Part II. Accuracy of the

estimates

Accuracy Examination:

EPS revenue

n% more accurate

Estimize error

Wall street error n

% more accurate

Estimize error

Wall street error

>=1 analyst 5251 51.20% 37.80% 49.50% 5691 48.10% 13% 14.70%

>=3 analysts 2734 52.30% 38.30% 57.30% 2979 48.50% 12.80% 15%

>=10 analysts 839 51.10% 24.70% 41% 918 48.90% 5.34% 4.68%

>=20 anaysts 248 53.20% 27.40% 28.80% 288 53.50% 6.40% 5.45%

Sector More accurate

Information Technology 60.50%Consumer Staples 62.60%

Telecommunication Services 52.90%Utilities 46%Industrials 58%Materials 51.80%

Consumer Discretionary 60.90%Financials 57.50%Health Care 59.50%Energy 51.70%

Accuracy of Estimize – Sorted by Sector:

Part III. Earnings surprise

0 1 1 2 2 n ndaily return X X X residual return

Some Concepts:

Earnings surprise:(actual EPS – estimated EPS)/ estimated EPS

Residual return:

Size: total assets Value: P/EGrowth: revenue growth last year Leverage: D/EMomentum: trailing 1 year return Yield: Dividend yieldIndustry: dummy variableVolatility: Standard deviation of trailing 1 year daily returns

Larger residual returns will begenerated for the trading days afterthe earnings surprise, if theestimates are more accurate

Rationale:

Despite the impact of incomplete data resulted fromdefunct stock tickers and missing historical values,estimates from Estimize still beat the those from Wallstreet when considering the earnings surprise effect .

Part IV. Deviation from the benchmark

Delta ─ A measure of deviation

Definition: Percent discrepancy between the Estimizeestimation and the Wall Street estimation in the daysleading up to the report date.

Rationale:

Institutional investors trade onnumbers provided by the sell-side. Estimize delta shouldprovide an early indication ofsuch tradings. We look at thecumulative daily residual returnafter a 10% or larger delta.

The cumulative event returns as a function of thetrading days after a significantly large delta ---- thepredictive power of Estimize estimation is welldemonstrated from this angle.

Delta effect by Market capital:

• The time span we use may be different fromthe exact time period that the prior study hasused.

•Defunct tickers and missing historical data mayjeopardize the integrity of the input.

• The vagueness in the regression method usedin the initial study brings challenge to ourreplication.

Potential Sources of Discrepancy

In spite of all the possible sources ofdiscrepancy, the “wisdom of crowd”effect is still pretty significant !

Thank you for listening !

Conclusion