This work is about data fusion and multi-fidelity data in finance. Data fusion has recently been widely used in finance data analyses. Portfolio optimization, a special multi-objective optimization problem, is rebalancing asset allocation. This work aims to improve portfolio optimization using multi-fidelity data and data fusion methods to generate a new portfolio, which can bring higher returns and a better Sharpe ratio. Our work is the first attempt to apply multi-fidelity for portfolio optimization. We define fidelity in portfolio optimization. We propose two different multi-fidelity models in portfolio optimization. One is the sampling interval based multi-fidelity model. The other one is the prediction based multi-fidelity model. According to the experiment results, the sampling interval based multi-fidelity model is feasible. We can generate a fused long-term portfolio from short-term and long-term portfolios in the sampling interval based multi-fidelity model. This fused long-term portfolio can outperform long-term portfolios. Moreover, this model can refine the Sharpe ratio and the return by a binary prediction of local optimal gamma.