I am a beginner in time series forecasting using ML, and I am encountering a strange phenomenon. I have air quality data, in which I have information of various pollutants. The goal is to predict AIR QUALITY INDEX.
There are about 10 pollutants and various weather features. The AQI feature(ground truth) has been created from the dataset itself(ignore why this is so), but note that two features, PM10 and PM2.5 are highly correlated to target AQI once it is created(0.8 & 0.81 respectively).Also note that this means that AQI at time(t) is being created from pollutant & weather information at time(t). For 1 step forecasting, I have shifted AQI column 1 step vertically - so now it becomes a problem where ML model must learn to predict 1 day future AQI using data available at time t.
I used random forest for forecasting. Also note that I created additional features such as lag of AQI by 1 day, 2 day ,rolling 30 day and 7 day average. I have not used rolling mean or standard deviation of PM2.5 & PM10.
The strange phenomenon: The forecasts are lagging.
I know this maybe due to the underlying process being random walk , but I highly doubt so. The actual AQI for 6 years look like that. Any ideas why this might be happening?