Real-world datasets for portfolio selection and …

1 Data ArticleReal-world datasets for portfolio selectionand solutions of some stochastic dominanceportfolio modelsRenato Brunia, Francesco Cesaroneb,n, Andrea Scozzaric,Fabio TardelladaDip. Di Ingegneria Informatica, Automatica e Gestionale, Sapienza Universit Di Roma, Rome, ItalybDip. di Studi Aziendali, Universit di Roma Tre, Rome, ItalycFacolt di Economia, Universit degli Studi Niccol Cusano, Rome, ItalydDip. Metodi e Modelli per l'Economia, il Territorio e la Finanza, Sapienza Universit di Roma, Rome, Italyarticle infoArticle history:Received 26 February 2016 Received in revised form9 June 2016 Accepted 21 June 2016 available online 28 June 2016abstractA large number of portfolio selection models have appeared in theliterature since the pioneering work of Markowitz.

2 However, evenwhen computational and empirical results are described, they areoften hard to replicate and compare due to the unavailability of thedatasets used in the Real-world price values from several major stock markets. Thedatasets contain weekly return values, adjusted for dividends and forstock splits, which are cleaned from errors as much as possible. Thedatasets are available in different formats, and can be used as bench-marks for testing the performances of portfolio selection models andfor comparing the efficiency of the algorithms used to solve them. Wealso provide, for these datasets , the portfolios obtained by severalselection strategies based on Stochastic Dominance models (see OnExact and Approximate Stochastic Dominance Strategies for PortfolioSelection (Bruni et al.))

3 [2])). We believe that testing portfolio modelson publicly available datasets greatly simplifies the comparison of thedifferent portfolio selection The Authors. Published by Elsevier Inc. This is an open accessarticleundertheCCBY license( ).Contents lists available atScienceDirectjournal in The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license( ).nCorresponding author. Phone: 39 06 Cesarone).URL: in Brief 8 (2016) 858 862 Specifications TableSubject areaEconomics and FinanceMore specific sub-ject areaPortfolio selection , portfolio optimization, Asset allocationType of dataTables, textfiles, excelfiles, matlabfiles,figuresHow data wasacquiredThomson Reuters Datastream, Fama & French Data LibraryData formatProcessed,filtered, analyzedExperimental factorsWhen necessary.

4 The assets prices arefiltered to check and to correct missing orinaccurate dataExperimentalfeaturesAll data sets provided consist of weekly assets returns readily usable in PortfolioSelection modelsData source locationN/AData accessibilityData is within this articleValue of the data The datasets provided here can be used as benchmarks by researchers willing to implement and tocompare portfolio selection models on publicly available data. If different researchers use the same publicly available data, the comparison of different approa-ches would be more easy and fair. The data arefiltered to remove possible errors in the original source. This allows researchers toperform more accurate and realistic simulations and evaluations. For our datasets we also provide the solutions to several portfolio selection models.

5 Such solutionscan be used by other researchers to compare the efficiency of their algorithms and the quality oftheir solutions. Availability of data and solutions can stimulate contacts among researchers working in this area forfuture collaborations and DataWe provide weekly returns time series for assets and indexes belonging to several major stockmarkets across the world. Weekly returns data are computed from prices values obtained fromThomson Reuters Datastream( ) and from daily returns obtainedfromFama & French Data Library( ). The data arefiltered to check and to correct missing or inaccurate values. The dataprovided can be used as input for several types of portfolio selection models to compare on bothefficiency and performance (for references on portfolio selection approaches see, ,[3]).

6 For theabove datasets , we also include as benchmarks the portfolios obtained by using several selectionstrategies based on both exact and approximate Stochastic Dominance models (described in[2]).2. Experimental design, materials and methodsAsset allocation aims at selecting a portfolio overNavailable assets in an investment universeA f1;..;Ngaccording to specific choice criteria under uncertainty. More precisely, we must decidehow much of each assetiAAshould be purchased in the selected portfolio . The portfolio is denoted byx fx1;..;xNg, wherexiis the fraction of the given capital invested in Bruni et al. / Data in Brief 8 (2016) 858 862859 Letpi;tdenote the price of assetiat timet, observed form 1 time periods, ,tA0;1;..; linear return of assetiat timetisri;t pi;t pi;t 1 =pi;t 1wheretAT 1.

7 ;m. Denoting bybtthe value of the benchmark ( , the Market Index) at timetA0;1;..;m, the benchmark linear returns arerIt bt bt 1 =bt 1wheretAT 1;..;m. The portfolio linear return at timetATisRt x XiAAxiri;tAll the datasets listed in the followingTable 1containjTj mlinear return values for each of theNassets contained in the market, together with the linear returns of the benchmark index, computed asdescribed 1 5 consist of weekly linear returns computed on daily price data, adjusted for dividendsand stock splits, obtained fromThomson Reuters Datastream. The selected benchmark is the marketindex. Stocks with less than ten years of observations were disregarded, thus obtaining a reasonabletradeoff between the number of assets (N) and of observations (|T|).

8 Furthermore, when necessary, theassets prices arefiltered to check and to correct inaccurate data. Data cleaning is indeed an importantissue for similar data (see, ,[4]for references on this widespread problem).Dataset 6 is derived from the Fama and French 49 Industry portfolios, available from the Fama &French Data Library, which contains daily returns from July 1926 to July 2015. Since there are manydata missing, especially before July 1969, we choose a subsample ofH 11628 periods where all theTable 1 Weekly returns datasets Name#of assets(N)jTjTime intervalCountry Description#of rebalancing(nreb)1 DowJones281363 Feb 1990-Apr2016 USADow Jones IndustrialAverage1102 NASDAQ100 82596 Nov 2004-Apr2016 USANASDAQ 100463 FTSE10083717 Jul 2002-Apr2016 UKFTSE 100564SP500442595 Nov 2004-Apr2016 USAS&P 500465 NASDAQComp 1203685 Feb 2003-Apr2016 USANASDAQ Composite536FF49 Industries 492325 Jul 1969-Jul 2015 USAFama and French 49 Industry190 Table 2 portfolio selection models applied to the NameDescriptionCZeSDCumulative Zero-order epsilon Stochastic Dominance (see[1,2])

9 RMZ_SSDR oman-Mitra-Zviarovich Second-Order Stochastic Dominance(see[9])LR_ASSDL izyayev-Ruszczynski approximate Second-Order StochasticDominance (see[5])L_SSDL uedtke Second-Order Stochastic Dominance (see[6])KP_SSDPost-Kopa Second-Order Stochastic Dominance (see[8])MeanVarMarkowitz Mean-Variance (see[7])R. Bruni et al. / Data in Brief 8 (2016) 858 862860daily returns of the 49 industries are available , namely from July 1969 to July 2015. Furthermore, tostandardize the frequencies of all data sets we extract weekly returnsrwkby cumulating daily returnsrdiin groups offive as follows:rwk 5j 11 rd5k j 1;k 0;..;H5 1:Since no market index is publicly available for the Fama and French 49 Industry portfolios, in thiscase we use the Equally-Weighted portfolio as a benchmark addition to the returns datasets , we also make available the composition (weights) and the out-of-sample returns of the portfolios obtained, for all datasets and for several in-sample periods, withthe models listed inTable 2and fully described in the companion paper[2].

10 For each dataset and for each model, we compute the solutions using a rolling in-sample windowof 52 returns observations. We initially set the in-sample window on thefirst 52 time periods, weselect the portfolio by solving the model, and we evaluate the performance of the selected portfolioon the following 12 (out-of-sample) periods. Next, we update the in-sample window, with theinclusion of the previous 12 out-of-sample periods and the exclusion of thefirst 12 periods of theprevious in-sample window. We then rebalance the portfolio by solving the model again, and repeatuntil the end of the dataset (seeFig. 1).Following the notation ofTable 1, the data provided with this article are organized as inFig. 2andlabeled as follows: : matlab workspace containing thejTjXNreturns matrix (Assets_Returns) and thejTjX1 vector of Index returns (Index_Returns) for theDataset.