Introduction to factorselect

Introduction

The factorselect package implements six estimators for determining the number of factors in large dimensional approximate factor models. The estimators differ in their theoretical assumptions, computational approach, and finite sample performance.

The recommended estimator for most applications is the Ahn and Horenstein (2013) eigenvalue ratio estimator, which is robust to perturbations in the eigenvalue spectrum and performs well when only one of N or T is large.

Simulating Factor Model Data

The package includes a helper function for simulating data from a static approximate factor model:

\[X = F \Lambda' + E\]

where \(F\) is a \(T \times k\) matrix of factors, \(\Lambda\) is an \(N \times k\) matrix of loadings, and \(E\) is an \(N \times T\) matrix of idiosyncratic errors.

set.seed(42)
X <- simulate_factor_model(N = 100, TT = 200, k = 3, sd = 0.5)
dim(X)
#> [1] 200 100

The Recommended Estimator: Ahn & Horenstein (2013)

The eigenvalue ratio (ER) and growth ratio (GR) estimators of Ahn and Horenstein (2013) are obtained by maximizing the ratio of adjacent eigenvalues of the sample covariance matrix. The ratio approach provides robustness to perturbations in the eigenvalue spectrum.

A key advantage over Bai and Ng (2002) is that the Ahn-Horenstein estimator works well when only one of N or T is large, not requiring both dimensions to grow simultaneously.

result <- select_factors(X, method = "ahn_horenstein", kmax = 8)
print(result)
#> Factor Number Selection
#> =======================
#> Call: select_factors(X = X, method = "ahn_horenstein", kmax = 8)
#> 
#> kmax: 8 
#> 
#> Estimated number of factors:
#>   ahn_horenstein       3

Comparing All Estimators

All six estimators can be run simultaneously by passing a vector of method names:

result_all <- select_factors(
  X,
  method = c("ahn_horenstein", "bai_ng", "abc",
             "lam_yao", "onatski_2009", "onatski_2010"),
  kmax   = 8
)
print(result_all)
#> Factor Number Selection
#> =======================
#> Call: select_factors(X = X, method = c("ahn_horenstein", "bai_ng", 
#>     "abc", "lam_yao", "onatski_2009", "onatski_2010"), kmax = 8)
#> 
#> kmax: 8 
#> 
#> Estimated number of factors:
#>   ahn_horenstein       3
#>   bai_ng               3
#>   abc                  3
#>   lam_yao              6
#>   onatski_2009         3
#>   onatski_2010         3

Scree Plot

The plot method produces a scree plot of the leading eigenvalues with the selected number of factors marked for each estimator:

result_ah <- select_factors(X, method = "ahn_horenstein", kmax = 8)
plot(result_ah, main = "Scree Plot — Ahn & Horenstein (2013)")

Finite Sample Performance

To illustrate the finite sample performance of the estimators, we run a small simulation study with 100 replications across three sample size configurations.

set.seed(123)
n_reps  <- 100
k_true  <- 3
configs <- list(
  large_both  = list(N = 100, TT = 200),
  small_N     = list(N = 25,  TT = 200),
  small_T     = list(N = 200, TT = 25)
)

results <- lapply(configs, function(cfg) {
  estimates <- replicate(n_reps, {
    X <- simulate_factor_model(N = cfg$N, TT = cfg$TT,
                               k = k_true, sd = 0.5)
    res <- select_factors(X,
                          method = c("ahn_horenstein", "bai_ng",
                                     "onatski_2010"),
                          kmax   = 8)
    res$k
  })
  rowMeans(estimates == k_true)
})

# Percentage correct for each configuration
do.call(rbind, lapply(names(results), function(nm) {
  data.frame(
    config         = nm,
    ahn_horenstein = round(results[[nm]]["ahn_horenstein"] * 100),
    bai_ng         = round(results[[nm]]["bai_ng"] * 100),
    onatski_2010   = round(results[[nm]]["onatski_2010"] * 100)
  )
}))
#>                     config ahn_horenstein bai_ng onatski_2010
#> ahn_horenstein  large_both            100    100           87
#> ahn_horenstein1    small_N            100    100           60
#> ahn_horenstein2    small_T            100      0           96

The simulation confirms that Ahn and Horenstein (2013) performs well across all three configurations, including when only one dimension is large. Bai and Ng (2002) tends to be less reliable in the asymmetric sample size cases.

Notes on Individual Estimators

Bai & Ng (2002) and ABC (2010)

These estimators use unstandardized data internally. The select_factors function handles this automatically — users do not need to preprocess data differently when requesting these methods.

Lam & Yao (2012)

This estimator uses lagged auto-covariance matrices rather than the contemporaneous covariance matrix. The number of lags h defaults to 1 but can be adjusted:

result_ly <- select_factors(X, method = "lam_yao", kmax = 8, h = 1)
print(result_ly)
#> Factor Number Selection
#> =======================
#> Call: select_factors(X = X, method = "lam_yao", kmax = 8, h = 1)
#> 
#> kmax: 8 
#> 
#> Estimated number of factors:
#>   lam_yao              6

Onatski (2009)

This estimator performs a sequential hypothesis test. The significance level alpha defaults to 0.05 but can be adjusted:

result_o09 <- select_factors(X, method = "onatski_2009",
                              kmax = 8, alpha = 0.05)
print(result_o09)
#> Factor Number Selection
#> =======================
#> Call: select_factors(X = X, method = "onatski_2009", kmax = 8, alpha = 0.05)
#> 
#> kmax: 8 
#> 
#> Estimated number of factors:
#>   onatski_2009         3

Onatski (2010)

The edge distribution estimator uses an iterative calibration procedure to estimate the threshold separating systematic from idiosyncratic eigenvalues:

result_o10 <- select_factors(X, method = "onatski_2010", kmax = 8)
print(result_o10)
#> Factor Number Selection
#> =======================
#> Call: select_factors(X = X, method = "onatski_2010", kmax = 8)
#> 
#> kmax: 8 
#> 
#> Estimated number of factors:
#>   onatski_2010         3

References

Ahn, S.C. and Horenstein, A.R. (2013). Eigenvalue Ratio Test for the Number of Factors. Econometrica, 81(3), 1203-1227.

Bai, J. and Ng, S. (2002). Determining the Number of Factors in Approximate Factor Models. Econometrica, 70(1), 191-221.

Alessi, L., Barigozzi, M. and Capasso, M. (2010). Improved Penalization for Determining the Number of Factors in Approximate Factor Models. Statistics and Probability Letters, 80, 1806-1813.

Lam, C. and Yao, Q. (2012). Factor Modelling for High-Dimensional Time Series: Inference for the Number of Factors. The Annals of Statistics, 40(2), 694-726.

Onatski, A. (2009). Testing Hypotheses About the Number of Factors in Large Factor Models. Econometrica, 77(5), 1447-1479.

Onatski, A. (2010). Determining the Number of Factors From Empirical Distribution of Eigenvalues. The Review of Economics and Statistics, 92(4), 1004-1016.