This thesis work studies a robust empirical regularity in currency exchange market, known as Uncovered Interest Rate Parity ( UIP ) Puzzle, that currencies with higher interest rate tends to appreciate in the short run. This thesis presents a multi-agent, multi-good general equilibrium model to explain the puzzle, inspired by external habit formation model of Campbell and Cochrane (1999). The model has four features, (i) the model delivers stochastic risk premiums, (ii) the external habit in the model is defined over goods rather than countries, (iii) the effect of external habit on different agents is asymmetric, (iv) the model has closed form solutions. The first feature is inherited from the standard habit formation models, and is one key ingredient of resolving the UIP puzzle as pointed out by Fama (1984). The second feature is different from the standard external habit formation models, which usually define external habit on country level. The good level external habits act as global risk factors that drive the dynamics of the pricing kernels in different countries. The third feature is generated by assigning different habit preference parameters to different countries. Asymmetric loading is claimed by Backus, etc. (2001) to be a crucial feature for affine term structure models to explain UIP puzzle. Combined with the second feature, this setting can also account for the puzzle raised by Engel(2012) and claimed to be difficult to explain with existing models, that is the long-run currency move ments are exactly the opposite relative to the short-run movements. With the forth feature of the model, one can conduct rigours econometric estimation in addition to calibration, which is a common exercise in habit formation literature. With consumption, yield curve, price level and exchange rate data for U.S., J.P. and U.K., the model is estimated using Markov Chain Monte Carlo method. The estimated model delivers a close fit of the term structures of interest rate for three countries, UIP puzzle and upward sloping real yield curves simultaneously. With the estimated parameters, the Engel's puzzle can be matched locally.