آموزش کارت امتیازی متوازن هیئت مدیره به منظور بهبود عملکرد شرکت

The objective of this paper is to demonstrate how the boosting approach can be used to define a data-driven board Balanced Scorecard (BSC) with applications to S&P 500 companies. Using Adaboost, we can generate alternating decision trees (ADTs) that explain the relationship between corporate governance variables, and firm performance. We also propose an algorithm to build a representative ADT based on cross-validation experiments. The representative ADT selects the most important indicators for the board BSC. As a final result, we propose a partially automated strategic planning system combining Adaboost with the board BSC for board-level or investment decisions.

Kaplan and Norton [20] introduced the Balanced Scorecard (BSC) as a management system that helps organizations define their vision and strategy, and translate them into specific actions. The BSC provides feedback on internal business processes, performance, and market conditions in order to review the strategy and future plans [21], [22], [23], [24] and [28]. Large U.S. companies, such as General Electric and Federal Express, and non-profit and public organizations have implemented the BSC approach [2] and [36]. The strategy of an organization, its main objectives, and its key business drivers define the indicators of the BSC. However, the choice of indicators is, in general, highly subjective and is often driven by company management or industry practices. Youngblood and Collins [39] describe a method based on indicators using multi-attribute utility theory. Clinton et al. [6] base their method on Analytic Hierarchy Process; nevertheless, these methods still require a mix of quantitative measures with a qualitative evaluation by managers or experts. The main objective of this paper is to adapt a machine learning method, such as Adaboost, to define the core variables and the structure of the board BSC. The criterion used to design the board BSC is the firm performance. We compare the predictive capacity of Adaboost with several other algorithms such as logistic regression, and other decision trees. The rest of the paper is organized as follows: Section 2 presents the basic concepts of a board BSC; Section 3 presents the methods used in this paper; Section 4 introduces the data and variables used in this research; Section 5 explains in detail our experiments; Section 6 presents the results of our forecast; Section 7 examines the results and the transformation of a representative ADT to a board BSC, and Section 8 presents the conclusions.

Boosting has been applied to different business areas such as direct marketing [27], financial forecasting [1], and electronic commerce [35]. In these areas, the analysis may substantially improve if boosting is used not only for prediction but also for interpretation as this paper has demonstrated. Adaboost, combined with the representative ADT, is an algorithm that can partially automate the definition of a board BSC as we proposed in our initial hypothesis. This algorithm is able to forecast corporate performance, select the most important variables, establish relationships among these variables, define a target for each variable to optimize corporate performance, and build a board strategy map and a board BSC. With this tool, managers can concentrate on the most important strategic issues and delegate the calculation of the targets to a semi-automated planning system supported by Adaboost. The use of ADTs in finance requires time-series or cross-sectional data in order to calculate meaningful nodes. Indicators that do not have enough information cannot be quantified using ADTs, so, the initial versions of a board BSC still require an important participation of the board of directors, middle and senior management. However, as the planning team or the company creates its own database, then the representative ADT can select the relevant indicators and their targets. As Creamer and Freund [8] showed, Adaboost also worked adequately with small datasets. However, the variance of the test error increased as the size of the dataset decreased. We suggest that companies that use Adaboost to build board BSCs use large datasets (industrial surveys or compensation surveys) or build their own internal dataset using the company's historical information. Finally, this paper can be enriched by the modification of the representative ADT algorithm to reduce the loss of accuracy in relation to the average test errors of the original ADTs. This difference is related to the variance of the test errors of the ADTs that also depends of the size of the dataset used. Future research can be directed to optimize the rules that define when and where a node is included in the representative ADT. For instance, we could combine several splitter nodes with the same predecessor at the same level. We could also have a more complex representative ADT that might be different from a single ADT, although may simulate more closely the diverse sample of ADTs. When there is a small training dataset, then generating an ADT with a small number of decision nodes is useful both in terms of interpretability and in terms of performance on the test set. However, when there is a large training dataset it is necessary to make a choice: either performance or interpretability. A small ADT will be easy to interpret and to translate into a BSC with an emphasis on understanding how the different variables interact as we have done in this paper. A large ADT will be more appropriate if the scorecard is used for very precise calculations such as those required in risk management or forecasting.