Search

Using a series of advanced tests from White's (2000) “Reality Check” to correct for data-snooping bias, we assess the out-of-sample performance of various portfolio strategies relative to the naive 1/N rule. When we analyze 16 basic portfolio strategies, 126 learning strategies, and nearly 2,000 extended strategies, we find that some strategies outperform the 1/N rule in conventional tests that do not account for data-snooping bias. However, after we use the new tests that control for such bias, we find that none or very few of these strategies outperform the 1/N rule. Thus, our finding underscores the necessity to control for data-snooping bias when making asset allocation decisions.

We propose a "1/N favorability index" to measure how favorable a market is to holding a 1/N portfolio. This index reflects the extent of difficulty for an optimized portfolio to outperform the 1/N portfolio in a specific market. A single-factor model predicts that bull markets are accompanied by a high 1/N favorability index and vice versa. We validate the model implication that the 1/N portfolio is more difficult to beat in bull markets using stock return datasets from a number of countries as well as the classic datasets used by DeMiguel et al. (2009). Our results imply that the reported good performance of the 1/N portfolio in the US equity market can be partially attributed to the long-run bullish trend in the market which gives rise to the high favorability of the market to the 1/N portfolio.

Universities looking to recruit or to rank researchers have to attribute credit scores to their academic publications. While they could use indexes, there remains the difficulty of coauthored papers. It is unfair to count an n-authored paper as one paper for each coauthor, i.e., as n papers added to the total: this is “feeding the multitude” . Sharing the credit among coauthors by percentages or by simply dividing by n (“1/ n rule”) is fairer but somewhat harsh. Accordingly, we propose to take into account the productivity gains of parallelization by introducing a parallelization bonus that multiplies the credit allocated to each coauthor. It might be an idea for coauthors to indicate how they organized their work in producing the paper. However, they might systematically bias their answers. Fortunately, the number of parallel tasks is bounded by the number of coauthors because of specialization and the credit is bounded by a limiting Pareto maximum. Thus, the credit is given by ( N + 2 ) / 3 n for N parallel tasks. As there may be, at most, as many parallel tasks as co-authors, credit allocated to each coauthor is given by ( n + 2 ) / 3 n , that varies between 2/3 of a single-authored paper for two coauthors and 1/3 when the number of coauthors is very large. This is the “maximum parallelization credit” rule that we propose to apply. This new approach is feasible. It can be applied to past and present papers regardless of the agreement of publishing houses. It is fair and it rewards genuine cooperation in academic publishing.

Using a series of tests based on White's (2000) "Reality Check," we assess the out-of-sample performance of a large class of portfolio strategies relative to the naive 1/N rule and control for data-snooping bias. We examine if 23 basic strategies and more than 5,000 extended strategies outperform the 1/N rule in terms of the Sharpe ratio and the certainty-equivalent return (CEQ). Our empirical results based on various stock portfolios suggest that only a few strategies out-perform the 1/N rule after we control for data-snooping bias and transaction costs. Thus, this paper has implications for market efficiency and asset allocation.

This paper provides an analysis of the behavior of separate-and-conquer or covering rule learning algorithms by visualizing their evaluation metrics and their dynamics in coverage space, a variant of ROC space. Our results show that most commonly used metrics, including accuracy, weighted relative accuracy, entropy, and Gini index, are equivalent to one of two fundamental prototypes: precision, which tries to optimize the area under the ROC curve for unknown costs, and a cost-weighted difference between covered positive and negative examples, which tries to find the optimal point under known or assumed costs. We also show that a straightforward generalization of the m-estimate trades off these two prototypes. Furthermore, our results show that stopping and filtering criteria like CN2's significance test focus on identifying significant deviations from random classification, which does not necessarily avoid overfitting. We also identify a problem with Foil's MDL-based encoding length restriction, which proves to be largely equivalent to a variable threshold on the recall of the rule. In general, we interpret these results as evidence that, contrary to common conception, pre-pruning heuristics are not very well understood and deserve more investigation.

It is unfair to count a n-authored paper as one paper for each coauthor, i.e., as n papers: this is “feeding the multitude”. Sharing the credit among coauthors by percentages or by simply dividing by n is fairer but somewhat harsh. So, we propose to take into account the productivity gains of parallelization by introducing a team bonus function that multiplies the allocation thereby increasing the credit allocated to each coauthor.The degree of parallelization cannot be determined exogenously discipline by discipline. So, one may propose that each team of coauthors indicates how the labor was organized to produce the paper. Unfortunately, the coauthors may systematically bias their answers in order to increase the team bonus by three mechanisms: (i) the coauthors exaggerate the number of parallel tasks, (ii) the team exaggerates the number of coauthors, (iii) the team exaggerates the parallelization rate.Fortunately, (i) the number of parallel tasks is bounded by the number of coauthors; (ii) cheating about the number of coauthors reduces the reward that each receives; (iii) we show that the reward is bounded by a limiting Pareto-optimal case. Thus, the bonus is given by (N 2)/3 for N parallel tasks; when we have the minimum number n of coauthors, the same number of coauthors as parallel tasks, the reward is given by (n 2)/3n, that varies between 2/3 of a paper for two coauthors and 1/3 of a paper, asymptotically.This new approach is feasible, and fair and it rewards genuine cooperation in academic publishing.

4N Baird's rule represented the extension of Hückel's 4N + 2 rule to triplet state systems. In this work we extend the 2(N + 1)(2) Hirsch rule for spherical aromatic species to open-shell spherical compounds and we provide evidence that those spherical species having a same-spin half-filled last energy level with the rest of the levels being fully-filled, i.e., those having 2N(2) + 2N + 1 electrons and S = N + ½, are aromatic.

In this paper, we wish to evaluate the performance of simple asset-allocation strategies such as allocating 1/N to each of the N assets available. To do this, we compare the out-of-sample performance of such simple allocation rules to about ten models of optimal asset-allocation (including both static and dynamic models) for ten data sets. We find that the simple assetallocation rule of 1/N is not very inefficient. In fact, it performs quite well out-of-sample: it typically has a higher Sharpe ratio, a higher certainty equivalent value, and a lower turnover than the policies from the optimal asset allocation. The intuition for the good performance of the 1/N policy is that the loss from naive rather than optimal diversification is smaller than the loss arising from having to optimize using moments that have been estimated with error. Simulations show that the performance of policies from optimizing models relative to the 1/N rule improves with the length of the estimation window (which reduces estimation error) and also with N (which increases the gains from optimal diversification). But, even with an estimation window of 50 years, the difference in the performance of the 1/N policy and the policies from models of optimal asset allocation is not statistically significant.

An m-of-n concept is a restricted form of the disjunctive normal form. The purpose of this paper is firstly to introduce a learning algorithm that induces m-of-n rules, and secondly to present the performance of this algorithm when applied to the DNA domain. This paper explores a search method for m-of-n concepts that gives rise to a rule induction algorithm called MoN. A pruning strategy akin to alpha-beta pruning is employed to reduce the search space encountered in the exhaustive search. A set of rules that consists of a disjunction of m-of-n concepts is used in MoN, as opposed to the disjunction of conjunctive rules normally used in rule induction algorithms. We show the application of this algorithm in inducing rules for primate splice-junction gene sequences (DNA) and compare the results obtained from C4.5 rules and ID2-of-3 concepts. >