The U.S. Federal Reserve Board goes to great lengths to make quarterly forecasts of GDPas accurate as possible, as these forecasts play a key role in informing monetary policy that
affects millions of Americans. It is therefore not surprising that there is ongoing effort to
refine and redefine existing GDP forecasting models so they are more accurate, especially
in the event of random shocks to the economy. During the COVID-19 shock, few economic
models were able to capture the full volatility of the GDP swings resulting from a once-in-
a-lifetime global pandemic. Since then, a whole slew of models, from MIDAS regression to
MF-VAR, have been used to forecast the economic impacts of COVID-19 more realistically.
In this thesis, I propose a new approach – applying Principal Component Analysis (PCA)
to a mix of monthly and quarterly economic variables – to forecast GDP per quarter from
1960 to 2023. To account for the mixed frequency nature of the data, I build PCA forecasts
across three data organizations: 1) using only quarterly data, 2) using stacked data, and 3)
using imputed data and select the best fit model. I then assess performance of my model
against the most widely used times series methods. I find that my economic data is highly
sensitive to overfitting and that the PCA model using only quarterly data performs the best
at predicting GDP, including capturing the COVID-19 shocks.