Estimating reliable transfer pricing profit indicators requires use of all available data

By Ednaldo Silva, Founder & Director at RoyaltyStat

Consider two aspects of statistical reliability principles. First, reliability can be measured by the ratio of the selected variable estimate divided by its standard error.  We want this reliability ratio to be as high as possible. Second, reliability depends on sample size such that larger samples produce more reliable estimates.

Transfer pricing rules recognize that statistical estimates of the selected profit indicator must consider multiple-year analysis to achieve a reliable measure of arm’s length taxable income. See OECD (2017), ¶¶ 3.75-3.79 and US Treas. Reg. § 1.482-1(f)(2)(iii).

OECD (2017), ¶ 3.76 states:

“In order to obtain a complete understanding of the facts and circumstances surrounding the controlled transaction, it generally might be useful to examine data from both the year under examination and prior years.”

OECD (2017),  ¶ 3.38 also states:

“The identification of potential comparables has to be made with the objective of finding the most reliable data, recognising that they will not always be perfect.”

In practice, though, multiple-year analysis is reduced to the audit year plus two years before or after the audit year.

Here, we show by one actual (as opposed to imagined) example that multiple-year analysis must consider all the available years of data of each comparable. Otherwise, estimates of the selected comparable profit indicator are not as reliable as the available data permit.

This conclusion is supported by OECD (2017), ¶ 3.57, which states:

“[I]f the range includes a sizeable number of observations, statistical tools that take account of central tendency to narrow the range (e.g. the interquartile range or other percentiles) might help to enhance the reliability of the analysis.” Emphases added.

We aim to satisfy two concurrent objectives. First, we need to stabilize the estimate of the selected profit indicator, avoiding wobbly numbers. Second, we aim to achieve the maximum reliability of the estimate.

We are confident that regression analysis is the most appropriate statistical algorithm (contrary to rote computing of quartiles) to satisfy these two objectives. Now we wish to show that more years of comparable company (enterprise) data are better than using truncated three-year data samples.

We estimate the operating profit markup model and examine the behavior of the estimated markup and its standard error, measured by the t-statistics (which is the ratio of the OLS (ordinary least squares) coefficient estimate divided by its standard error).

     (1)     S(t) = α + β C(t) + U(t)

     (2)     U(t) is corrected for serial correlation by the Newey-West algorithm,

and t = 1 to T years of data.

The variable S(t) denotes Net Sales of the selected comparable enterprise in period t, and C(t) denotes Total Costs (Lato) = (COGS + XSGA + (DP − AM)). Kalecki (1954), Chapter 1 (Cost and prices), is the most original reference to the profit markup model.

The residual errors U(t) have zero mean and Newey-West corrected variance. See Green (2018), §20.5, pp. 996-999. Wikipedia has a good entry on the Newey-West estimator of the standard errors of the OLS regression coefficients at this link

RoyaltyStat employs both Gauss and R-Package algorithms; they produce similar results. See OECD (2017), ¶¶ 3.57-3.62 for a discussion of statistical tools. 

Since OLS requires a certain minimum degree of freedom (sample size minus two observations in the bivariate case of model (1)), we start with = 3 years of data considering the most recent to previous values. As an exception, the t-statistics = 168.3 for = 3 years of data are from OLS because the number from Newey-West is a large outlier resulting from its standard error being near zero.

The (S(t), C(t)) data are from RoyaltyStat’s integration of the Capital IQ (Compustat) database of listed global company financials. We use Best Buy (GVKEY 2184) data from 1983 to 2018 and apply OLS to our bivariate regression model (1):

  t = 3, β = 1.0813, t-statistics = 168.3             t = 21, β = 1.0431, t-statistics = 171.2

  t = 4, β = 1.1214, t-statistics =   51.8             t = 22, β = 1.0450, t-statistics = 181.7

  t = 5, β = 1.1312, t-statistics =   34.0             t = 23, β = 1.0468, t-statistics = 186.4

  t = 6, β = 0.9857, t-statistics =   10.2             t = 24, β = 1.0477, t-statistics = 195.0

  t = 7, β = 0.9373, t-statistics =   26.4             t = 25, β = 1.0479, t-statistics = 207.3

  t = 8, β = 1.0499, t-statistics =   33.9             t = 26, β = 1.0479, t-statistics = 219.4

  t = 9, β = 1.0703, t-statistics =   45.3             t = 27, β = 1.0478, t-statistics = 230.1

t = 10, β = 1.0786, t-statistics =   55.3             t = 28, β = 1.0477, t-statistics = 239.2

t = 11, β = 1.0793, t-statistics =   60.0             t = 29, β = 1.0475, t-statistics = 246.8

t = 12, β = 1.0623, t-statistics =   49.8             t = 30, β = 1.0474, t-statistics = 253.3

t = 13, β = 1.0416, t-statistics =   40.7             t = 31, β = 1.0473, t-statistics = 258.7

t = 14, β = 1.0326, t-statistics =   55.8             t = 32, β = 1.0472, t-statistics = 263.5

t = 15, β = 1.0313, t-statistics =   78.0             t = 33, β = 1.0471, t-statistics = 267.6

t = 16, β = 1.0316, t-statistics =   99.1             t = 34, β = 1.0470, t-statistics = 271.4

t = 17, β = 1.0347, t-statistics = 117.8             t = 35, β = 1.0470, t-statistics = 274.7

t = 18, β = 1.0364, t-statistics = 132.6             t = 36, β = 1.0469, t-statistics = 277.6

t = 19, β = 1.0396, t-statistics = 142.5

t = 20, β = 1.0413, t-statistics = 157.1

To induce rule change, we make a compelling case for the use of all available comparable data in transfer pricing analysis.

The convergence of the estimated profit markup coefficients to a stable value and their accompanying t-statistics as the comparable company sample size gets larger, and the scatterplots with the superimposed Loess curves, provide strong support to our statistical principle and fact-based position that reliable estimates of arm’s length taxable income require that we use all available data.

Restricting the data to three-year company samples in practice today produces unreliable (unstable) estimates of the selected profit indicator together with inefficient standard errors (and thus very wide ranges).

REFERENCES

William Green, Econometric Analysis (8th edition), Pearson, 2018.

Michael Kalecki, Theory of Economic Dynamics, George Allen & Unwin, 1954. 

OECD, Transfer Pricing Guidelines (2017). DOI: https://dx.doi.org/10.1787/tpg-2017-en

Scatterplot of Coefficient vs Count

Scatterplot of Newey-West vs Count

 

Ednaldo Silva

Ednaldo Silva

Founder & Director at RoyaltyStat

Dr. Ednaldo Silva is Founder & Director of RoyaltyStat, a leading online database of royalty rates extracted from unredacted license agreements filed with the SEC.

He is an economist with over 25 years of experience in transfer pricing innovation and the valuation of intangibles.

Dr. Silva helped draft the US transfer pricing regulations as Senior Economic Adviser in the IRS Office of Chief Counsel. He was the originator and developer of the “comparable profits method” and introduced the best method rule and the concept that arm’s length is represented by a range of results. Dr. Silva was also the first economist in the IRS's Advance Pricing Agreement (APA) Program.

Ednaldo Silva
Ednaldo Silva
Managing Director
RoyaltyStat LLC

6931 Arlington Road, Suite 580 | Bethesda, MD 20814-5284 | USA
Telephone 1-202-558-2356 | http://www.royaltystat.com

Be the first to comment

Leave a Reply

Your email address will not be published.