Long/Short Sector-based - Duke's Fuqua School of Business



Long/Short Sector-based

Trading Strategy

Emergent Asset Management, LLC

Presented by:

Konstantin Savov

Scott Smith

Pin-Yew Wong

Vaswar Mitra

Vinaya Jain

Table of Contents

Project summary 3

Project Methodology 4

Objective and strategy outline 4

Description of factors 4

Scoring methodology 5

Considerations in choosing sectors and final choice 6

Sector-wise detailed results 7

Financials 7

Average annual factor returns for top and bottom quintiles 7

Factors selected and calculated scores 7

Factor correlations 8

In sample analysis 8

Other Results 9

Out of sample analysis 9

Healthcare 10

Average annual factor returns for top and bottom quintiles 10

Factors selected and calculated scores 10

Factor correlations 11

In sample analysis 11

Other results 12

Out of sample analysis 12

Industrials 13

Average annual factor returns for top and bottom quintiles 13

Factors selected and calculated scores 13

Factor correlations 14

In sample analysis 14

Other results 15

Out of sample analysis 15

Model imperfections and further research 16

Project summary

Our project aims to identify factors that are successful at predicting future returns for stocks within a particular industry sector. We aim to develop a quantitative stock screening model that can be used to rank stocks within a sector universe into different fractiles. The principal objectives of our analysis are:

1) To identify factors (accounting, expectational, momentum or market sentiment based), which have predictive power of sector returns using a top down screening process.

2) Use these factors to build a combined scoring system that can be used to rank the individual equities within a particular sector. This scoring system would be generated by looking at sector returns within an “in sample” period of data.

3) To implement quantitative long/short market neutral equity strategy based on this scoring system. After we use the combined score to rank the stocks into fractiles, we would go long the first fractile of equities and short the last fractile. We would also evaluate whether such a strategy delivers any outperformance over the benchmark or alpha by testing it over a period of “out of sample” market data.

Project Methodology

Objective and strategy outline

Our objective was to develop a quantitative long/short sector-based stock selection model that generates positive and consistent results. This project tested the hypothesis that it is possible to generate positive returns using a top down approach even in arguably efficient markets. We took this analysis a step further and analyzed how different factors apply for different sectors and whether diverse sectors can generate similar alphas.

We defined our universe using the S&P1500 index and chose stocks with a market capitalization greater than five hundred million. The S&P 1500 index includes large, mid and small cap stocks in the US and combines the S&P 500, S&P 400 (Mid Cap Index) and the S&P 600 index of small cap stocks. Our sector choices were based on S&P’s GICS industry classification system. Due to data availability problems in the years prior to 1996, we limited our sample period from 1996 to 2003. The portfolio was rebalanced monthly.

We selected an initial list of thirteen factors and performed a sector wise analysis to establish the most suitable factors for each sector. We applied our own objective scoring methodology to determine the optimal weights for each of these factors in the final combined score screen in the final combined scoring screen. We then reran the alpha tester to check whether our choice of factors and weights showed a trend in returns. The out of sample period for our analysis was 2004 and 2005. We went long the first quintile and short the fifth quintile and checked again whether our out of sample period showed the same promising results as the in sample period. We employed an equal weight strategy, both in the number of stocks 'per bucket' and in the amount invested or shorted per stock. Although our benchmark is value-weighted, we thought equal weight was appropriate for simplicity and to limit the impact from outlier performers.

Description of factors

We defined factors from each of the three categories: Fundamental, Expectational and Momentum. Our alpha tester tested the predictability of each of these factors and we chose the most suitable ones for a given sector.

Fundamental

1. Book to Price: book value per share / price per share

2. Dividend Yield: dividends per share / price per share

3. CFO Yield: cash flow from operations / price per share

4. Earnings Yield: LTM earnings / price per share

5. Return on Assets: annual net earnings / total assets

6. % Change in ROA: % change in ROA over previous month

7. Return on Equity: annual net earnings / total shareholder equity

8. % Change in ROE: % change in ROE over previous month

Expectational

1. Revision Ratio: (Upward revisions – downward revisions) / total revisions

2. SUE Score: standard unexpected earnings (Earnings surprise / std deviation)

3. Mean FY1 to Actual % Change: (FY1 Estimate – Current Actual) Current Actual

4. LT Projected Growth – Historical Growth: Next 5 year annualized estimate – previous 5 year annualized actual

Momentum

1. 12-Month Lagged Monthly Price Growth: Eg: (Jan’06 Price – Jan’05 Price) / Jan’05 Price

Scoring methodology

We created an objective scoring methodology of our own that is explained here using an example.

[pic]

1. In each month the quintiles are ranked based on returns. In the example above, on 5/1/2003, quintile 1 (F1) had the highest return and quintile 5 (F5) had the second lowest return.

2. Points are assigned from highest to lowest rank (+2, +1, 0, -1, -2). Pursuing the same example, F1 gets +2 which F5 gets -1, based on their respective ranks.

3. The total points for the in sample period (05/2003 to 12/2003) are summed for the two quintiles.

4. Steps 1 through 3 are repeated for each factor.

5. The raw point scores are scaled across factors from +10 to -10.

6. However not all the scored parameters (the ten scores shown above) are considered for the scoring screen. This scoring methodology objectively rewards consistency of results. Based on this objective method and on our subjective evaluation of return trends as well as factor correlations, we chose only some of these factors for our final scoring screen.

Considerations in choosing sectors and final choice

We limited our universe based on the S&P1500 index and only considered stocks that had a market cap higher than five hundred million. We felt that this was a good balance between having a sufficiently large number of companies and avoiding some of the unnecessary volatility that comes with small cap stocks. This range would also give us sufficient liquidity necessary to execute the monthly rebalancing strategy. Furthermore, we applied the sector restriction using S&P’s GICS classification system. We started with five sectors: Financials, Healthcare, Industrials, Energy and Information Technology (IT) and later pared the list down to just the first three. The energy sector did not have sufficient number of companies in our defined universe. The IT sector has grown rapidly during the last few years and has been excessively volatile. This instability manifested itself in a lack of any predictability of returns with respect to our factors. We therefore chose the first three sectors which had sufficiently large number of companies in the defined universe and showed some predictability with respect to the chosen factors.

Sector-wise detailed results

Financials

Average annual factor returns for top and bottom quintiles

[pic]

Factors selected and calculated scores

[pic]

Factor correlations

[pic]

In sample analysis

Heat map of multivariate factor returns

[pic]

Average multivariate factor returns across quintiles

[pic]

Other Results

[pic]

Out of sample analysis

Heat map of multivariate factor returns

[pic]

Average multivariate factor returns across quintiles

[pic]

Healthcare

Average annual factor returns for top and bottom quintiles

[pic]

Factors selected and calculated scores

[pic]

Factor correlations

[pic]

In sample analysis

Heat map of multivariate factor returns

[pic]

Average multivariate factor returns across quintiles

[pic]

Other results

[pic]

Out of sample analysis

Heat map of multivariate factor returns

[pic]

Average multivariate factor returns across quintiles

[pic]

Industrials

Average annual factor returns for top and bottom quintiles

[pic]

Factors selected and calculated scores

[pic]

Factor correlations

[pic]

In sample analysis

Heat map of multivariate factor returns

[pic]

Average multivariate factor returns across quintiles

[pic]

Other results

[pic]

Out of sample analysis

Heat map of multivariate factor returns

[pic]

Average multivariate factor returns across quintiles

[pic]

Model imperfections and further research

Given the time constraints, there are several imperfections in our analysis and represent areas for further investigation and potential improvement of results. Firstly, our portfolios were rebalanced monthly without any consideration of the trading costs involved. We did not consider potential restrictions or difficulties of executing short sales, particularly in the small cap universe. There are several options to improve this analysis. Firstly we could evaluate the trade off between the cost of turnover and the benefit of rebalancing frequently. This analysis might reveal that rebalancing quarterly or semi-annually is a better strategy than monthly rebalancing. Another way might be to combine this top-down approach with a bottom-up approach and ensure that fundamentally solid companies are retained in the long portfolio even if they temporarily fall off the top quintile. The same goes for fundamentally poor companies which should be retained in the short portfolio, even though they may show short term positive interfractile migration. Another approach to evaluate individual companies is through interfractile migration analysis. This analysis would tell us to include stocks that have remained in the top quintile through most of their history but have fallen off the list temporarily and to exclude unstable stocks that have a history of drifting across quintiles and therefore lead to excessive portfolio turnover.

The second limitation of this model is the subjectivity of the scores even though we tried to incorporate a certain level of objectivity into this process. A potential area for improvement is to allow the optimizer to select long and short weights for a portfolio of quintile 1 and quintile 5 factor returns. The optimizer would however ignore the impact of a shift in trends over time. The problem is that the weights are static and we can potentially resolve this problem through dynamic weights. One way to do this might be to shift the in-sample time horizon with the passage of time. For example the in-sample time horizon in Jan 2006 would be Jan 1994 to Dec 2003. In Jan 2007 the horizon would shift a year and would be Jan 1995 to Dec 2004. Another approach for dynamic weighting would be through interaction variables. For example an interaction variable for the sign of yield curve slope would yield different factor weights depending on the positive or negative sign of the yield curve slope.

Several of the approaches mentioned above can be used in conjunction to develop a more sophisticated model.

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

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

Google Online Preview   Download