Beneish M-Score Backtest

Are we in a post-truth era? Fake news, lies and misstatements dominate our headlines. What’s a value investor to do? There are a few tools in our fundamental analysis toolkit to tackle this issue. One such metric is the Beneish M-Score.

Messod D. Beneish published a paper back in 1999 titled, The Detection of Earnings Manipulation. In that paper, Beneish details a probit model for detecting earnings manipulators. The model is composed of eight key factors based on fundamentals. The resulting score from this model is referred to as the Beneish M-Score. The model identifies earnings manipulators when the M-Score exceeds -1.78. A higher numerical M-Score scores are associated with increased probability of manipulation. The M-Score is based on the following formula:

M-Score = -4.84 + 0.92*DSRI + 0.528*GMI + 0.404*AQI + 0.892*SGI +0.115*DEPI – 0.172*SGAI – 0.327*LVGI + 4.679*TATA

Where:
Days Sales in Receivables Index (DSRI) = (Net Receivables_t / Sales_t) / (Net Receivables_t-1 / Sales_t-1)
Gross Margin Index (GMI) = [(Sales_t-1 – COGS_t-1) / Sales_t-1] / [(Sales_t – COGS_t) / Sales_t]
Asset Quality Index (AQI) = [1 – (Current Assetst + PP&Et + Securitiest) / Total Assetst] / [1 – ((Current Assetst-1 + PP&Et-1 + Securitiest-1) / Total Assetst-1)]
Sales Growth Index (SGI) = Sales_t / Sales_t-1
Depreciation Index (DEPI) = (Depreciation_t-1/ (PP&E_t-1 + Depreciation_t-1)) / (Depreciation_t / (PP&E_t + Depreciation_t))
Sales General and Administrative Expenses Index (SGAI) = (SG&A Expense_t / Sales_t) / (SG&A Expense_t-1 / Sales_t-1)
Leverage Index (LVGI) = LVGI = [(Current Liabilitiest + Total Long Term Debtt) / Total Assetst] / [(Current Liabilitiest-1 + Total Long Term Debtt-1) / Total Assetst-1]
Total Accruals to Total Assets (TATA) = (Income from Continuing Operations_t – Cash Flows from Operations_t) / Total Assets_t

The Portfolio123 version of this uses the depreciation component of Dep&Am by looking at the annual depreciation values that Compustat makes available and calculating a ratio to extract Depreciation from Dep&Amort interim data.

I ran several backtests of the Beneish M-Score on Portfolio123 using my standard backtest rules. I did add the financial sector using the rule “Sector != FINANCIAL” and I also excluded stocks that didn’t have a Beneish M-Score in Portfolio123. I ran the backtest from July 1, 2000 to February 1, 2021 using 52 week holding periods. I divided the stocks into 5 equal quintiles by their M-Score rank from smallest to largest. I also ran a separate backtest for M-Score above the -1.78 cutoff recommended to identify potential earnings manipulators.

Summary of Results for Beneish M-Score

* Average excess returns were analyzed starting in each month with 12-month holding periods (256 sample periods). This avoids the potential for seasonal reporting bias.

As you can see from the chart above, the first three quintiles of stocks ranked by M-Score don’t really display much of a difference in average excess returns with a one-year holding period. Those lower scores actually indicate there is a lower likelihood that those companies manipulated earnings. What’s important here is that the top 20% of stocks ranked by M-Score in the fifth quintile actually do display an almost 3 percent lower average annual excess earnings. The median M-Score for the fifth quintile is only -1.89. If you then just test stocks with M-Scores above -1.78, average annual excess returns then drop to -4.54% for the test period (July 1, 2000 to February 1, 2021). Below is a chart of just the returns for the stocks rebalanced annually that have M-Scores above -1.78 that might indicate earnings manipulation.

As you can see from this chart, high M-Score seems to result in consistently lower returns over the test period. The M-Score might be an advanced ratio worth considering to help filter out potentially higher risk, lower performing companies. I recommend backtesting your own strategies on Portfolio123 with this M-Score as an added filter to select out potential earnings manipulators.