Instructions:

1-for the first part: In fact, it is a matter of doing the same thing twice: once with the forecast error as the dependent variable, and again with the forecast dispersion as the dependent variable. The first step is to show that, all other things being equal, the earnings forecast errors are higher in times of crisis (2020)... which is very likely, but we still have to show it. for this, we focus on the Covid crisis. The aim is to show that the forecast errors for the year 2020 are greater than the forecast errors for previous years. We take all the forecasts issued by all the analysts, calculate the forecast error and explain the forecast error by the variable of interest, control variables and fixed effects. - The dependent variable is the forecast error normalized by the stock price. - The variable of interest is a dummy variable that we can call Crisis, which is worth 1 if the forecast (and its error) is relative to 2020. It is 0 otherwise. The control variables are:

- Ln of the firm's market capitalization

- Loss: dummy = 1 if the reported EPS is negative

- Ln (1+Horizon): number of days between the date of the forecast and the date of announcement of the forecasted result

- Ln(1+Nfirm) : the number of firms followed by the analyst - Experience or Ln(1+Experience): the seniority of the analyst in years For the fixed effects, we suggest sector effects (SICs), or even country effects (NCOU). We expect the coefficient associated with Crisis to be significant if the forecasting errors of the results are higher during a crisis. Beware, forecast errors can be positive or negative. It is therefore necessary to work on the absolute value of the forecast errors, then separately on the negative and positive errors. So 3 regressions (we shall work on sub-periods also).

2- In a second step, we show that, all things being equal, the forecasting errors during the crisis are lower when the analyst is experienced and highly specialized. We use the same model as before without the Crisis variable, focusing only on forecast errors for the year 2020. The two variables of interest are NFirm or Ln(1+Nfirm) for specialization and Experience or Ln(1+Experience) for experience. If the assumptions hold, the coefficient associated with Experience or Ln(1+Experience) must be significant and negative (meaning that when an analyst is experienced, he doesn't commit too much errors), the coefficient associated with NFirm or Ln(1+Nfirm) must be significant and positive. We will also try to do regression by forecast horizon for 2020 (0-90 days , ...) I already did some computations in the excel worksheet. I am not an expert in regression (since i never studied econometrics) but i think that the data set is unbalanced.

First part: In fact, it is a matter of doing the same thing twice: once with the forecast error as the dependent variable, and again with the forecast dispersion as the dependent variable. The first step is to show that, all other things being equal, the earnings forecast errors are higher in times of crisis (2020)... which is very likely, but we still have to show it. for this, we focus on the Covid crisis. The aim is to show that the forecast errors for the year 2020 are greater than the forecast errors for previous years. We take all the forecasts issued by all the analysts, calculate the forecast error and explain the forecast error by the variable of interest, control variables and fixed effects. - The dependent variable is the FORECAST ERROR (normalized by the stock price). - The variable of interest is a dummy variable that we can call CRIS, which is worth 1 if the forecast (and its error) is relative to 2020 forecast. It is 0 otherwise. The control variables are:

- Log(CAP) : Logarithm of the firm's market capitalization

- LOSS: dummy = 1 if the reported EPS is negative

- Ln (1+HORZ): number of days between the date of the forecast and the date of announcement of the forecasted result

- Ln(1+COMP) : the number of firms followed by the analyst

- Ln(1+EXP) : the seniority of the analyst in years For the fixed effects, I suggest sector effects (SICs), or even country effects (NCOU).

- Relative FORECAST ERROR We expect the coefficient associated with CRIS to be significant if the forecasting errors of the results are higher during a crisis (we will compare between 2020 and 2018-2019). Beware, forecast errors can be positive or negative. It is therefore necessary to work on the absolute value of the forecast errors, then separately on the negative and positive errors. So :

- First: for this method in a bulk.

- Then: do the same thing with sub-periods . Part 2 : We use the same model as before without the Crisis variable, focusing only on forecast errors for the year 2020. The two variables of interest are ln(comp) or Ln(exp) for specialization and Experience + the same control variables (except ln(comp) and ln(exp) which are independent variables here). If the assumptions hold, the coefficient associated with Ln(1+Exp) must be significant and negative (meaning that when an analyst is experienced, he doesn't commit too much errors), the coefficient associated with Ln(1+comp) must be significant and positive. ? We will work with sub periods.

1.Subject or field:

Earnings forecast accuracy and Covid-19.

2.Explanation:

FIRST PART:

Prior literature focused on the quality of forecasts, thus, the measure of forecasts' error or accuracy - one being the opposite of the other forecast_error = (actual_value - predicted_value) / stock market price.

Dispersion is also analyzed: at a given moment, the idea is to see if the consensus is strong or weak, i.e. whether we have concentrated forecasts or dispersed ones. At a given moment, we will have several analysts who will issue forecasts, and we will see if they are highly concentrated or dispersed. The idea is to say that if:

-they are very dispersed: the consensus is weak. The analysts do not share the same vision of the firm. There is a lot of dispersion, so there is a lot of asymmetry of information - everyone has their own vision of the firm.

In the prior literature, it was modelized as the inter-analyst standard deviation of forecasts deflated by stock price.

Therefore, in our first part of the study, we will focus on measuring forecasts' errors and dispersion. We will try to compare the accuracy and dispersion between 2020 and benchmark periods (2018 and 2019). We will try to see if in 2020, all things being equal, there are greater or lower forecast errors / dispersion.

The study/regression is going to be conducted firm-by-firm and analyst-by-analyst (we will not measure average/mean errors).

Methodology:

When working on forecast errors, the question is: at what date shall we look at the forecast error?

Pertaining to my research question, we must look at the forecast error from the moment when everyone was aware of the Covid crisis. So, we have to determine important dates: I propose the dates of the different confinements for instance, date of the first lockdown: end of March

When analysts published their 2019 forecasts around February/March 2020 : the first forecast we're interested in is the first forecast after the earnings announcement.

After that, our 2nd important date is the date when the market became aware of the Covid crisis (end of March).

We will look at the forecast just after the market became aware of the Covid crisis. We will look at the forecasts that were issued from March 15.

The summer was calm. But, it degenerated in September 2020 in all countries with a 2nd lockdown.

I propose to make two models:

1- In the first model: we will take all the forecast errors for all the firms and analysts (individually, not on average) and we will run the regression. We will introduce fixed effects :

-analyst: we have analyst codes in the database.

-and firm: we have the firm codes in the database: CUSIP, SEDOL or ISIN.

In this model, we will introduce a control variable in order to discriminate recent forecasts (close to the 2020 earnings announcements): the number of days between the forecast and the next earnings announcement date. It is important to remember that forecasting the day before the results differs from the forecast made 6 months before the announcement of the results. Here, we run the whole thing in bulk.

2- In the second model: we will run the same model without mixing all the forecasts. We will work with sub-periods:

-the first forecast after the 2019 earnings announcement.

-And after that, we'll take the forecasts just at the beginning of the covid crisis.

-...

-and we'll end with the last available forecast before the 2020 earnings announcement.

We're going to notice that there are a lot of forecasts right before the earnings announcement. In other words, at the end of 2020 and beginning of 2021, we will have plenty of forecasts on the 2020 earnings.

For European companies, we will work on annual results (for our study, we will focus on the fiscal year 2020).

SECOND PART:

Our second part of the study will be about explaining forecast errors with some variables. So, the second thing is to focus only on 2000's forecast errors and explain why some analysts have bigger forecast errors than others.

Among these factors, we will focus on two variables of interest:

-Analyst's experience: it will be measured by the seniority in the database. We will look at the first time in which the analyst appears in the database via its code. If he first appears in 2000, he will have 21 years of experience. We will try to query the online database so that it gives me the first date in which the analyst appears (Excel file : Detail history 1997 to 2021).

-Competition between analysts: measured by the analyst's following (number of analysts who follow the firm).

We will also have some control variables (firm-specific - specific to the difficulty of forecasting - and analyst specific). We are going to take the control variables used by Guido Bolliger (2004) see below : relative forecast accuracy, analysts' general forecasting skills,company-specific experience, portfolio complexity, geographical diversification, forecasts staleness.

8 / 27 + 89% 3.2. Measurement of dependent and independent variables The definition of variables are similar to those used in previous research. There- fore, I only provide the reader with a short description of each variable and place in brackets the study that first suggested it. PMAFE (Clement, 1999). The ratio of the current year individual analyst's fore- cast error for a particular firm divided by the mean current year forecast error of all analysts for the firm, minus one. It measures relative forecast accuracy GEXP (Lim, 2001). The logarithm of one plus the number of consecutive years (up to the current year) during which the analyst has been supplying at least one forecast. It measures analysts' general forecasting skills. CEXP (Lim, 2001). The logarithm of one plus the number of consecutive years (up to the current year) during which the analyst has been supplying at least one forecast for a specific firm. It measures company-specific (task-specific) expe rience . NCOMP (Clement, 1999). The number of firms for which the analyst supplied at least one forecast during the current year. It measures portfolio complexity. NSIC (Clement, 1999). The number of two-digit SICs for which the analyst sup- plied at least one forecast during the current year. This variable measures industry specialisation. NCOU (Clement et al., 2000). The number of two-digit I/B/E/S country codes for which the analyst supplied at least one forecast during the current year. It mea- sures analysts portfolio's geographical diversification. TOP10 (Clement, 1999). A dummy variable set to one if the analyst is employed by a firm ranked in the top 10% during the current year and set to zero otherwise. Brokerage houses are ranked yearly with respect to the number of analysts em- ployed. In the multiple-country regressions, TOP10 is computed by ranking bro- kerage houses on an European basis whereas for the country regression they are ranked on a country basis. It proxies for the resources available to the financial analyst.PMAFEijt = 1 . DCEXPijt + B2 . DGEXPijt + B3 . DNCOMPijt + BA . DNCOU;j,: + Bs . DNSICits + Bo . DTOP10jj + B, . DFAGEijt + Ejt (1) with all variables firm-year mean adjusted (D stands for cross-sectionally centered). I do not include a constant term since the respective means have been subtracted from each variable. A positive (negative) value for the centered variable means that the forecast error or characteristic of analyst i for stock j is above (below) average on year