This thesis is a compilation of three essays that bridge the theoretical and empirical study of financial markets. The subjects of study in the three main chapters are (i) equilibrium models of asset prices and asset holdings and trade; (ii) limited computational capacity and its interaction with asset prices and trades. In chapter 1 (joint with Peter Bossaerts) we show that statistical improvements can be made on a traditional test of portfolio "efficiency." Testing portfolio efficiency is used in the practice of investment decisions as well as to test theoretical models of asset prices (CAPM and multifactor models). We propose a parametric family of tests of the efficiency of a portfolio in a market with a risk-free asset. All tests in the family compare the mean-variance ratio of the tested portfolio (benchmark) with that of a different portfolio (reference). We show that the power of a test in our proposed family depends on the correlation between the benchmark and the reference portfolio. This provides a way to improve the power of efficiency tests for a given sample, by choosing the appropriate test in this family. Chapter 2 (joint with Peter Bossaerts and William Zame), is a test of the theory of dynamically complete markets. In this work we compare prices and portfolio choices in complete and incomplete experimental financial markets. The incomplete-markets treatment differs from the complete-markets one in that we close one market, and announce, halfway through trading, which of three states will not occur. We find prices and allocations to be analogous across the two treatments, as predicted by theory. In particular, subjects' additional trading in the incomplete-markets treatment is such that the final allocations become indistinguishable from the complete-markets treatment. The results show that participants form rational expectations about retrade prices, which is a very strong finding. Chapter 3 (joint work with Peter Bossaerts and Jernej Copic) moves away from existing theoretical paradigms. It explores the implications of analyzing intellectual discovery as the solution of a nonincremental problem, outside the reach of traditional models of learning with updating. The experiment sets up a situation that is non-incremental and where Bayesian updating is not a sensible model. In this framework we find that communication is possible, and that a primitive code is good enough to achieve intellectual discovery, not discourage it.