Structural models of the accounting process and earnings

This thesis investigates how the statistical properties of earnings depend upon the accounting process. Two features of the accounting process--historical cost accruals and the periodic reporting of earnings--are modeled. Their implications for two statistical properties of earnings--the time-series of earnings and the relation of price and earnings--are derived and empirically tested.;Earnings is modeled as "properly measured" earnings plus measurement error. Permanent earnings, a traditional "properly measured" earnings variable, is defined in a setting of valuation under uncertainty. The time-series of permanent earnings and the relation of price and permanent earnings are derived.;Historical cost accruals are shown to yield measurement error that follows a moving average process with order equal to the life of the assets and liabilities of the firm minus one. This specification of the measurement error leads to a model of the time-series of earnings-to-price-plus-dividends ratios. Earnings-to-price-plus-dividend ratios means revert based upon the serial correlation of the measurement error. Estimation of the model provides evidence that the measurement error is positively serially correlated, as predicted, and that the magnitude of the serial correlation depends on the amount and term of the historical cost depreciation accrual. The model is used to forecast earnings; it improves on conventional models in terms of mean square forecast error.;Periodic reporting is characterized through a model of earnings as temporally aggregated (time-averaged) permanent earnings. Periodic reporting smooths the earnings time-series. A model with the change in earnings as a function of quarterly changes in price plus dividends from the prior two years is derived. A distinctive (triangular) pattern of coefficients on the changes in price plus dividends is predicted and found empirically. The model is used to forecast earnings, and it improves upon conventional models.;Both historical cost accruals and periodic reporting yield measurement error in earnings, which leads to bias in valuation models. A structural equations estimation provides estimates of the bias in levels and first differences valuation models. The bias is much larger for first differences models.