Assignment 04 .edu



Assignment 01 – Screening Based Stock Selection, Industry Comparisons

Professor Campbell R. Harvey

Table of Contents

1. Objective 4

2. Market Selection 4

3. Industry Selection 4

4. Time Period 5

5. Factor Selection 5

Factor 1: One month price momentum 5

Factor 2: Three month price momentum 5

Factor 3: Dividend yield 6

Factor 4: Rate of reinvestment 6

Factor 5: Consensus forecast earnings estimate revision ratio 6

Factor 6: Change in consensus FY1 estimates 6

Factor 7: Operating cash flow / Sales 6

Factor 8: Interest coverage (EBIT / Annual interest expense) 7

6. Findings 7

Industry: Oil and gas extraction (SIC code = 13xx) 7

Industry: Food and tobacco products (SIC code = 20xx + 21xx) 8

Industry: Electronic & other electric equipment (SIC code = 36xx) 9

Industry: Communication (SIC code = 48xx) 10

Industry: Depository institutions (SIC code = 60xx) 11

US Common Stocks 12

7. Analysis 12

8. Bi-variate Screens 12

Industry: Oil and gas extraction (SIC code = 13xx) 13

Industry: Depository institutions (SIC code = 60xx) 14

9. Optimizing Weights 15

Industry: Oil and gas extraction (SIC code = 13xx) 15

Industry: Depository institutions (SIC code = 60xx) 15

10. Closing Thoughts 16

11. Appendix A 17

12. Appendix B 17

Mutual Fund Evaluation 17

Asian Markets Evaluation 17

Objective

The main goal of our project is to explore and extend the findings from the “Stock Selection in Emerging Markets: Portfolio Strategies for Malaysia, Mexico and South Africa” (Harvey, et al.) into another area of interest. This idea was prompted by the recent availability of FactSet as a learning tool and the potential practical applications resulting. After a number of iterations, we found great interest in exploring applications segmented by industries within the US market, and the potential to optimize on these constraints.

Market Selection

In selecting a different area to explore based on discussions, we considered the mutual fund industry and other Asian countries beyond Malaysia. On the basis of data availability during initial trial, we opted for the more data-rich US market. An elaboration on why we thought these would be good projects is included in Appendix B, and perhaps worthy of another look given access to more information.

The refinement of our market selection was also very much influenced by the screening factors we chose, which will be explained in detail below. Specifically, we had found that the eight factors we chose produced very little tradable spreads in a US common stock universe within the historic in-sample time frame of 1990.01 – 2000.12.

In our discussion as to why these would vary in substance to the three countries considered in the paper, we hypothesized that perhaps the less developed nature of these countries were more dependent on some core industries, and that perhaps the US market does achieve noticeable spreads if we consider individual industries instead of the market as a whole, which is significantly more diversified. In addition, we saw value in exploring whether certain screening factors were industry specific, and how that would relate to the market as a whole.

Industry Selection

To standardize our selection of industries, we decided to base our study on Standard Industrial Classification (SIC) codes. This was reinforced by the fact that FactSet provides for the creation of universes by these codes. We evaluated each code based on a subjective assessment of the industry relevance to our study, as well as a dataset minimum of 100 companies. The minimum was set to allow us the ability to look for meaningful patterns within each industry without being adversely skewed by a disproportionate number of “N/A” listings for our factors.

The industries selected are listed below:

|Code |Industry |Dataset size (Companies) |

|13xx |Oil and gas extraction |199 |

|20xx + 21xx |Food & kindred and tobacco products |128 |

|36xx |Electronic & other electric equipment |506 |

|48xx |Communication |245 |

|60xx |Depository institutions |766 |

We recognize that both the companies listed within the SIC codes and the companies themselves can change over time. However, as new companies emerge and others falter, we believe this was a reasonable universe since the overall selection was still a representation of the particular industry. Some companies may change their main business over time as well by entering new markets or exiting their primary market. We felt that this impact would be minor overall.

Time Period

The aim was to balance the time period selected with the time it would take FactSet to run the factor screens. We decided on an 11 year period (1990.01 – 2000.12) as our in-sample run. The desire was to include a small measure of the recent recession to balance the exceptional run in the US market in the 1990s. This left 2001.01 – 2003.12 for our out-of-sample comparisons. Although this is a relatively short out-of-sample period, we believe that the results will allow us to gain insight into whether these factors are still relevant for our chosen industries. In hindsight, although the 11 year period provided a great data screen, it proved cumbersome to run. Most runs took between 20 to 45 minutes to complete.

Factor Selection

When picking screening factors we considered three criteria. First, we chose factors that had plenty of data available. Second, we chose at least two momentum, fundamental, and expectation factors. Third, we chose factors that performed well when used in predictive regression models. Note that in our analysis, we refer back to this as reference and generally refer to our factors by their designation number. Our eight main variables are:

Factor 1: One month price momentum

Data source: COMPUSTAT

Formula: G_PRICE_1MCHG

Reasoning: We chose one month momentum because we wanted to see if the contrarian nature of this factor held true over a range of industries. Additionally, momentum factors have performed well in past predictive regressions.

Factor 2: Three month price momentum

Data source: COMPUSTAT

Formula: G_PRICE_3MCGH

Reasoning: The three month momentum variable provides a nice comparison for the one month momentum results. We suspected that the one month momentum factor would lead to a superior long/short spread since that data set contained more information than the three month data. Additionally, momentum factors have performed well in past predictive regressions.

Factor 3: Dividend yield

Data source: COMPUSTAT

Formula: G_DIV_YLD(NR 0 L45D)

Reasoning: Strong dividend yields and dividend growth are consistently solid indicators of healthy companies.

Factor 4: Rate of reinvestment

Data source: COMPUSTAT

Formula: G_REINV_RATE(NR 0 L45D)

Reasoning: We believe this variable will capture the upside performance of the large, non-dividend paying firms. This variable also describes the general health of a firm. Firms with surplus cash and long term growth prospects tend to have greater reinvestment.

Factor 5: Consensus forecast earnings estimate revision ratio

Data source: COMPUSTAT

Formula: (IH_UP_FY1R(0)-IH_DOWN_FY1R(0))/IC_NEST_FY1R(0)

Reasoning: We are interested in factors that incorporated a change in values. We guess that firms with a greater number of upward (downward) revisions will over (under) perform.

Factor 6: Change in consensus FY1 estimates

Data source: COMPUSTAT

Formula: G_IH_MEAN_FYIR_3MCHG(0)

Reasoning: Again, we are interested in factors that incorporated a change in values. We guess that firms with greater per cent changes in earning estimates will benefit (suffer) from the effect of good (bad) news on the market

Factor 7: Operating cash flow / Sales

Data source: COMPUSTAT

Formula: G_CASH_FLOW(NR 0 L45D)/G_SALES(NR 0 L45D)

Reasoning: An interesting valuation ratio. We guess that companies with higher quality of sales will have superior returns.

Factor 8: Interest coverage (EBIT / Annual interest expense)

Data source: COMPUSTAT

Formula: G_INT_COV(NR 0 L45D)

Reasoning: An interesting valuation ratio. We looked for a variable that captured the idea of leverage. We guess that a firm with a lower coverage ratio will have lower returns since a larger portion of its capital is going towards interest commitments.

Findings

Upon evaluating our FactSet results (see MS Excel file A1_GTC.xls), we determined to focus on a few key factors and industries. Specific discussions are below. Note that some of the findings were difficult to rationalize.

Industry: Oil and gas extraction (SIC code = 13xx)

|Factor 1 |1 |2 |3 |4 |5 |

|SIC 13 Factor 1 Bucket 1 |SIC_13_F1_B1 |-0.1725 |0.0032 |0.0925 |0.0086 |

|SIC 13 Factor 1 Bucket 5 |SIC_13_F1_B5 |0.3738 |0.0226 |0.1125 |0.0127 |

|SIC 13 Factor 4 Bucket 1 |SIC_13_F4_B1 |0.8408 |0.0177 |0.0952 |0.0091 |

|SIC 13 Factor 4 Bucket 5 |SIC_13_F4_B5 |-0.0421 |-0.0008 |0.1105 |0.0122 |

|  |  |  |  |  |  |

|  |Portfolio |1.0000 |0.0228 |0.1000 |0.0100 |

|  |  |  |  |  |  |

|  |  |Weights |E |σUS-equity |  |

|  |Target |1.0000 |  |0.1000 |  |

Our annualized return is 27.3%. With no constraints on long or short positions, we were trying to set the portfolio σ = σUS-equity = 0.0454. However we found that the solution would not converge under the 0.1000 standard deviation we ended up with. We speculate that this is in large part due to the large volatility with these particular portfolios to start with. This is rationalized by the industry itself, which is very sensitive to economic factors. We are comforted by the duality of the weights for each of the two factors between their factors. From our discussion of the relative bucket merits, our long/short positions are as expected.

Industry: Depository institutions (SIC code = 60xx)

|Series |Code |Weights |E |Σ |σ2 |

|SIC 60 Factor 1 Bucket 1 |SIC_60_F1_B1 |-0.0923 |0.0037 |0.0580 |0.0034 |

|SIC 60 Factor 1 Bucket 5 |SIC_60_F1_B5 |0.9038 |0.0293 |0.0627 |0.0039 |

|SIC 60 Factor 2 Bucket 1 |SIC_60_F2_B1 |0.2438 |0.0053 |0.0572 |0.0033 |

|SIC 60 Factor 2 Bucket 5 |SIC_60_F2_B5 |-0.0554 |0.0253 |0.0726 |0.0053 |

|  |  |  |  |  |  |

|  |Portfolio |1.0000 |0.0260 |0.0600 |0.0036 |

|  |  |  |  |  |  |

|  |  |Weights |E |σUS-equity |  |

|  |Target |1.0000 |  |0.0600 |  |

Our annualized return is 31.2%. With no constraints on long or short positions, we were trying to set the portfolio σ = σUS-equity = 0.0454. However we found that the solution would not converge under the 0.0600 standard deviation we ended up with. While these four portfolios did not exhibit as high of volatility as in the oil & gas extraction industry, it was still significant enough to put a floor on what the resulting mean variance frontier stopped. Again, we are comforted by the duality of the weights for each of the two factors between their factors. From our discussion of the relative bucket merits, our long/short positions are as expected.

Closing Thoughts

This project focused on elaborating on what we developed as a screening exercise into areas relevant to the team. Driven to the US market due to data availability, we were quick to explore this methodology at the industry level. This would fit well into our top-down asset allocation from country and global asset class to a more tactical industry and firm selection and re-weighting.

The industry level exploration was interesting to us in comparison to how our factors performed against 1) the US market as a whole, 2) the three country equities selected in the paper that served as the basis for our project, and 3) each other.

The uni-variate screening allowed us to isolate industry specific relevant factors, while the bi-variate screening allowed us to take our best results and see if we could enhance the returns. We actually found that in certain situations, we actually lose the quality of information in going to a bi-variate screening due to how the independent factors “split” the other factors’ buckets. We also explored the optimization of scoring through the mean variance paradigm against a subjective scoring assignment in combining factor results.

In studying this element of asset allocation, we have been prompted with more ideas than we had time to process. Given resources and data, we would have liked to further explore:

▪ The multi-variate contrast between layering and a bi-variate screening, and how the systematic decoupling and re-coupling method compares with the independent.

▪ The predictive scope and range of the screening methodology with a multi-variate regression approach in achieving alphas, and within itself through bucket allocation methodologies.

▪ The emerging markets where efficiencies and regulation may derive returns (and risk) beyond that available within the US markets.

Our thanks to Prof. Harvey for his insight and feedback (and laptop), as well as Kevin Stoll for sharing a time-saving intellectual asset, without which we would have spent more time on data processing than on exploring new ideas.

Appendix A

Reference files:

▪ A1_GTC, Appendix A.PPS

▪ A1_GTC, Appendix A.MP3

Appendix B

Mutual Fund Evaluation

One of our first attempts was with the mutual fund sector. We hypothesized two main points. One, that we could create an optimizer that would allow individuals or firms to allocate an employees’ funds in an optimal way among a number of funds to create maximum returns based on a certain level of volatility. The second was that there may exist screens that would allow us to identify high performing funds among the many funds available.

The first idea was dropped because we were having some difficulty in getting to fund performance data quickly. Our unfamiliarity with Zephyr and our inability to create a universe of specific funds (Vanguard or Fidelity for example) in FactSet was problematic. As a result, we decided to attempt a general screen on all mutual funds within FactSet using a variety of factors. Although the number of funds was truly robust, we discovered another limiting factor. One of the elements that make screening equities so entertaining is that there are a seemingly infinite number of screening options at your disposal. Mutual funds however don’t have earnings reports, forward earnings estimates or other interesting metrics. Therefore, screens related to price seemed more appropriate. Rather than limit ourselves in this manor however, we decided to take a more stimulating run with equity screens.

Asian Markets Evaluation

Our second project idea involved constructing a long/short trading strategy for emerging equities markets in Asia. Our universe included approximately 14,000 equities from Korea, Taiwan, Singapore, Thailand, and six other high growth economies. We excluded the Japanese and Hong Kong markets, believing they were too mature. Our hypothesis was that by rebalancing our portfolio on a monthly basis we could exploit the inefficiencies and volatility of these markets and capture very large long/short spreads (30%-40%).

We used FactSet Alpha Tester to sort the returns based on a series of uni-variate screens. The screening variables included size, dividends, one month momentum, and E/P ratio as listed in the COMPUSTAT databases. The initial runs for years 1998-2002 were immediately disappointing. Due to lack of data, all but 40-80 companies were eliminated from the analysis. We then re-ran this analysis using the Japan and Hong Kong markets, but achieved no better results. We decided this small number of companies did not constitute a large enough sample and abandoned the project.

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download