Why Fama–French Portfolio Replications Change Across Data Vintages

# Mean of Excess Returns: Why Do Fama-French Library and F&F (2017) Differ for the Same Time Period and Portfolios?


# Mean of Excess Returns: Why Do Fama-French Library and F&F (2017) Differ for the Same Time Period and Portfolios?












I'm currently having a hard time finding the solution to my problem. I'm trying to replicate the 5 Factor Model from Fama & French (2017) for Europe. I'm using Data from the French Library and tried to calculate the Means of Excess Returns for the same time period 07.1990 - 12.2015. I also use the same Portfolio sortings. But my results differ from the paper. I calculated the Excess Returns by subtracting the riskfree rate (obtained from the Europe 5 factor Dataset) from the Portfolio Returns.

From what I've read, the Breakpoints for Portfolio construction are the same. Does someone know the answer? Or am I miscalculating something

Edit1:

Thanks for your reply, I'll try to be more precise. All Data I used is for Europe. I used the 5x5 Size, B/M Portfolio Dataset. So I had Returns on 25 Portfolios. With Data ranging from 07.1990 to 12.2015. The time period used in F&F (2017). Then I took the Riskfree rate from the "5 Factor for Europe" Dataset and subtracted it from the portfolio Returns. Month by Month for all 306 Datapoints. Thus I had the Excess Returns for all 25 Size, B/M Portfolios. Then I calculated the Mean of all 25 Excess Returns and compared it to the Paper. I did this in Excel and R. And in both cases there was difference in results of Means. For the means of the 5 factor Returns, I get pretty similar results. There are only very minor differences, which I think result from rounding differences. But I'm not able the calculate the same Excess Returns for the Portfolios. I also calculated the Regression and I only get approximately similar results to the paper. My supervisor told me to keep trying to get closer to the paper. But I really don't know how

Edit 2: Here's my R-Script for calculating the Means of Excess Returns:

> Returns <- read.csv("C:5x5_Size_BM.csv") Factor <- read.csv("C:/Europe_5_Factors.csv")

#Loading 5 Factor Dataset and Value-Weighted Portfolio Returns

> Factors <- Factor[1:306, 2:7] Returns <- Returns[1:306, 2:26]

#Removing datapoints after 12.2015 and first column with monthly dates

> Riskfree_Rate <- Factors[, 6]

#extracting the Riskfree rate data column into a single vector

> Riskfree_rate_Matrix <- matrix(rep(Riskfree_Rate, times = 25), nrow = 306, ncol = 25)

#turning vector into 25x306 matrix with identical datapoints

> Returns_Matrix <- as.matrix(Returns)

> Excess_Returns <- Returns - Riskfree_rate_Matrix

> Means_of_Excess_Returns <- colMeans(Excess_Returns)

> Means_Matrix <- matrix(Means_of_Excess_Returns, nrow = 5, ncol = 5, byrow = TRUE)

> print(Means_Matrix)

```
       [,1]      [,2]      [,3]      [,4]      [,5]
```

[1,] -0.1181046 0.2912418 0.3668301 0.5107190 0.6726471

[2,] 0.1800000 0.3951307 0.4839216 0.6519281 0.7114706

[3,] 0.3244771 0.5082026 0.4926471 0.5074837 0.6722549

[4,] 0.4579739 0.4795425 0.4864379 0.5061765 0.6013399

[5,] 0.3447712 0.5021569 0.5128431 0.6157190 0.5027451

## Answer by Enrico Schumann (score 2)

https://quant.stackexchange.com/a/81590

How close do you come with your results? Equity data are often revised (corporate actions are adjusted/corrected, ...). And IIRC, even on K. French's website, there is a warning that datasets can change. That being said, I just tried this (in R):

```
library("NMOF")  ## https://enricoschumann.net/R/packages/NMOF/index.htm
E5 <- French("~/Downloads/French",
             "Europe_5_Factors_CSV.zip")
i <- rownames(E5) >= "1990-07-31" & rownames(E5) <= "2015-12-31"
E5 <- E5[i, ]

E25 <- French("~/Downloads/French",
              "Europe_25_Portfolios_ME_BE-ME_CSV.zip")
i <- rownames(E25) >= "1990-07-31" & rownames(E25) <= "2015-12-31"
E25 <- E25[i, ]

E25.xreturns <- E25 - E5[, "RF"]

m.xreturns <- colMeans(E25.xreturns)
100 * matrix(m.xreturns, byrow = TRUE, nrow = 5)
##            [,1]      [,2]      [,3]      [,4]      [,5]
## [1,] -0.1181046 0.2915359 0.3668301 0.5102614 0.6728431
## [2,]  0.1800000 0.3950654 0.4839216 0.6519281 0.7114706
## [3,]  0.3162092 0.5082026 0.4927124 0.5071242 0.6724837
## [4,]  0.4588235 0.4795425 0.4864379 0.5075817 0.5997059
## [5,]  0.3447386 0.5021569 0.5128431 0.6157190 0.5027451
```

(Note that you'll need at least version 2.10-2 of the NMOF package to run the example.)

Comparing with the table on page 447, I get a maximum absolute difference of about 5 bp, which seems not too far off to me.

Update: I just fetched the two datasets from an archived version of Kenneth French's website, dated January 2017. When I redo the exact steps from above, I get the following matrix:

```
## with csv files as of Jan 2017
##            [,1]      [,2]      [,3]      [,4]      [,5]
## [1,] -0.1040523 0.2826144 0.4181699 0.4937255 0.6769608
## [2,]  0.2295425 0.3887255 0.4793464 0.6386275 0.7223529
## [3,]  0.2982353 0.5266993 0.5023203 0.5062418 0.6611765
## [4,]  0.4430719 0.4733660 0.4801634 0.5588235 0.6221242
## [5,]  0.3440196 0.4841503 0.5157516 0.6111438 0.4943464
```

Now all absolute differences are below 1 bp. (But of course there is no way of telling whether only the data changed, or the scripts that produced the data as well.)

## Answer by phdstudent (score 1)

https://quant.stackexchange.com/a/81593

It can indeed change over time. Here's a paper about it:

- Akey et al. Noisy Factors? The Retroactive Impact of Methodological Changes on the Fama-French Factors (2021)

And replicating FF factors can be very very difficult even for the U.S. What is your data source for the individual stock data?