I’ve added in the recession risk to test whether this would improve the forecast. The issue however is that my recession forecasts are at a monthly level vs. my forecasts are annual. To solve the issue in the interim, as soon as the monthly forecast signals a recession risk, I apply it to the whole year.
Figure 1: Monthly Forecast
Figure 2: Yearly Forecast
The issue with this is that it’s producing false signals. For example, the signal only triggered during the down period in March 2020 and the tail end of 2021. This caused false signals in between this range.
Another measure I checked was the % of forecasts that fell within the high/low forecast. With the recession plug, the accuracy dropped by over 50% in these periods.
Lastly, I looked at how the model forecasts the price for the S&P500 index. Without the recession plug, the 2008 recession period was significantly under forecasted and took time for it to catch up.
However, with the recession plug, it captured the dip quicker but still overshot the recovery period. It also significantly under forecasted for the 2020 period as the recovery was much swifter than a year.
In summary, the forecast will have to exclude the recession plugs for the time being. If the forecast were monthly, the recession model could work. But March 2020 taught me that government interventions can cause a quick recover at the expense of a potential recession in the future.
This confirms to me that the recession model needs a lot more work before I can have it out in production.