The objective of this project is to use neural networks to forecast next day's stock closing price. In the past, researchers used different methods to forecast stock price such as technical analysis, fundamental analysis, and economic analysis. Forecasting stock prices is a problem that has been usually approached in terms of weekly, monthly, or quarterly forecast. This project aims at finding a feasible way, by using neural networks, to make daily forecasts. Most methods proposed so far, such as technical, fundamental and economic analysis, are limited to solving the problem as a long term trend analysis. Thus, these methods either lack accuracy or add extra expenses to the forecasting task, especially if a company's fundamental statistics are out of date. Therefore it is difficult to forecast day-to-day close price as a nonlinear problem. In this study, three portfolios are created. Portfolio #1 is based on subjective forecasts, Portfolio #2 uses a neural network to forecast, and Portfolio #3 using CAPM optimizer forecast. A comparison of these portfolios showed that the CAPM optimization based on neural network forecast (Portfolio #3) achieved the highest return. The degree of accuracy is compared in three economic periods: the beginning of recession; the middle of recession; and the beginning of recovery. Stock forecasting example cases are given to illustrate this neural network approach to solve nonlinear problems. It is observed, indeed, that next day closing prices are forecast with better accuracy within a one-year period than other methods.