Model Construction

When constructing the different valuation strategies, we made several choices, which will be discussed in this section. We have pursued to rationalize all the choices by best practices from the literature and limit backtesting biases as much as possible.

The quantitative models have been constructed based on sound economic arguments rather than trying to optimize performance by fitting the variables (data mining). Instead, we present a total of 15 base valuation models and demonstrate robust performance across the board. We have pursued to test the impact of all modelling decisions throughout the thesis by stressing the assumptions of WACC, Growth, EBIT versus EBITDA, reported versus estimated FCFF etc.

The base models evaluate price/fair value based on the market cap at the closing price each month and invest at the same moment. In practice, it would not be possible to trade at the closing prices. This is one of the biases in our models. Yet, if we trade on the 1-month old price/fair values, the performance of the Gordon Growth models only diminishes marginally. Originally, Fama & French (1996, p. 61) found significant outperformance of the value factor even when lagging both financials and market capitalizations by 6 months. We have not tested whether our valuations are robust enough to outperform with a similar 6-month lag. It would depend on how fast market prices tend to mean-revert towards intrinsic value. Simultaneously, the effect of waiting before buying a stock that our models consider to be undervalued, does not necessarily have to be lower returns. If a stock appears undervalued because of negative price momentum, this could easily continue in the following months and impair our returns (Pedersen, Asness &

Moskowitz, 2013). Thus, waiting for any negative momentum to disappear before buying an undervalued stock could also potentially improve returns.

For the most important variables in the valuation models, such as growth and WACC, we evaluate the impact on both performance and sector exposure in the analysis. Some might consider this to be data mining, but the rationale have not been to arrive at the best performance, but rather to give the reader full visibility of the impact of changing critical variables. In the base model we chose to apply a sector specific WACC based on a sample test from Morningstar and a fixed growth rate of 3.95%. Morningstar's sector specific WACC turns out to be considerably lower than the WACC measures from Bloomberg and Damodaran (NYU), which provides a bias towards certain sectors and higher valuations. The low WACC from Morningstar results in more undervalued investments - thus more diversified portfolios and lower volatility. This combined with a relatively high growth rate to perpetuity of 3.95% further increase this effect. However, since the numerator of the quantitative valuations is somewhat more conservative, as we exclude the explicit forecast period, there is a limit to the impact of changes in WACC and growth.

Page 103 In the current environment of moderate growth and low inflation, it might seem aggressive to assume a long-term nominal growth rate of 3.95%. It is important to consider whether this high growth rate is reflected in the required return (WACC), since the spread between the two can otherwise become too narrow and result in unreasonably high valuations. The average WACC in our sector sample from Morningstar is 7.9%, the lowest is 6.4% (utilities), and the highest is 9%

(technology). This makes the spread between growth and required return around 3.5% to 5%. As clarified in Section 3, the steady-state growth rate should not exceed the growth rate of the economy, and for this reason, one might argue that 3.95% is too high (Damodaran, 2015 p. 40).

An important detail to consider in terms of free cash flows is the impact from acquisitions. The S&P 500 companies often make both small and large acquisitions, and consolidation has been a particularly popular way of growing since the financial crisis. When a company acquires another, FCFF might take a dramatic hit due to the cash out flow - depending on how the acquisition is financed. This can potentially decrease the attractiveness of a company in our valuation models, as we do not adjust the past financials for acquisitions. FCFF consists of cash flows from operating activities and investing activities. The outflow of cash from the acquisition will negatively affect investing activities, and if the acquisition is financed by issuing debt, the new debt will increase the cash flow from financing activities (not included in FCFF). All else equal, the result is a lower FCFF and a higher net interest bearing debt - both of which affects our Gordon Growth valuations negatively. In particular, the LY FCFF model could be prone to suffer from this effect. The effect of acquisitions on the value driver models should be less pronounced, as NOPAT is not directly exposed to the large outflow of cash, although returns on invested capital (ROIC) could be affected materially as goodwill and invested capital increases (McKinsey, 2015, p. 112).

The valuations of the value driver models depend directly on ROIC, which makes it essential that the assumptions to both operating profits and invested capital are realistic. We include both goodwill and intangible assets as invested capital, but the result of excluding these items would be a higher ROIC and higher valuations. This would also increase exposure towards sectors with a lot of acquired goodwill and intangible assets as a result of M&A activity or R&D. Examples of sectors with high returns on invested capital excluding goodwill would be technology, healthcare, and consumer staples (McKinsey, 2015, p. 108).

The backtest period from April 2003 to September 2018 has been particularly favorable for U.S.

stocks, and as a result, we find it important to discuss the consequences for our performance.

Firstly, the period includes two economic expansions with two associated bull-markets, and the bull-market since 2009 has been one of the longest in history. The period only showcases one major downturn during the financial crisis and a few moderate setbacks in 2011 and 2015.

Page 104 Secondly, the period does not include the Dot-com bubble that peaked in 2000, which potentially could impact the performance of our portfolios although most of them have not favored the relatively expensive technology sector. Yet, Morningstar’s 1-star rating had a large overweight in technology stocks in our 15-year backtest. After we had performed the backtest, the U.S. stock market experienced large monthly drops both in October and December 2018, which were not included in our measurement period. These could have influenced our results - especially those of the long-only portfolios but would also have affected the benchmarks.

The measurement period is relatively short compared to the other empirical studies of risk-factors presented throughout this thesis. This especially impacts the level of statistical significance of our long/short strategies and decile portfolios, as many of the result have insignificant t-values. In terms of our long only strategies the results look somewhat better in terms of t-values. We usually generate t-values above 1.96 which are significant at the 95%

level. We did not consistently find statistically significant long/short alphas above 1.96 (Pedersen, 2015, p. 29) and 3 (Harvey, Liu, and Zhu, 2015). However, both the level of alpha and t-values are generally slightly higher in the Gordon Growth models relative to the value driver models. We do not attribute the low t-values to a lack of performance in our backtest, but rather the short 15-year period. Statistical significance corresponds to realizing a high Sharpe ratio over a long period of time. Essentially, t-stat is approximately calculated as the Sharpe ratio multiplied by the square root of T (years) as described by Pedersen (2015, p. 53), so an otherwise attractive long/short Sharpe ratio of 0.40 would need roughly 60 years to achieve a t-stat of 3. As the Sharpe ratios are naturally higher on our long-only portfolios, this explains their higher t-stats.

It is best practice to test whether quantitative strategies work in other investment universes and markets over different time periods (Pedersen, 2015, p. 163). This increase the reliability of the backtest and probability of outperformance going forward. It has not been part of this project to test whether the strategies also worked on other markets in other periods of time, but it will definitely be something worth investigating in future research. We expect that when the strategy is implemented on broader markets with mid- and small-cap stocks as well, it will be positively affected by the size effect in line with other risk factors that also work better within small stocks (Alquist, Israel & Moskowitz, 2018, p. 8). Implementing the strategies on smaller stocks may also result in more losses from bankruptcies - especially for the Gordon Growth strategies, as they tend to favor firms with lower margins and higher financial leverage which indicates distress.

Page 105