Use ARIMA to forecast BTC price!
ACF and PACF graphs can help, here is how👇👇👇
The lag 0 (no lagged time series) is the original time series, so its correlation will always be 1. This lag is redundant and it can be ignored.
Values within the blue-shaded area are statistically non-different from 0 (no correlation).
Let's see how these graphs can help us get a first estimate of the parameters "p" and "q". 👇👇👇
But if you don't remember what ACF and PACF or lags mean, have a look at yesterday's thread first!
For the "p" parameter, which determines how many previous values (or prices in the BTC example) we need to consider for the AR part of ARIMA, we will check the PACF graph.
We can see two spikes out of the shaded area (apart from lag 0). Therefore, we should start with p = 2.
We can also see a spike out of the blue-shaded area in lag 12. However, this is too far from lag 0 and we don't want to have a very complex model with lots of lags to consider.
For the "q" parameter, which determines how many previous values of residuals we need to consider for the MA part of ARIMA, we will check the ACF graph.
We can see several spikes out of the shaded area. We will first consider the lags closer to 0, so can start with q = 3.
The 4th lag is not significant and the 5th and 6th lags, despite being significant, are further from the lag 0 and they are less significant.
However, we should check later if q = 6 is better than q = 3.
The reason why we are interested mainly in the lags close to the lag 0 is because we don't want a very complex model, we want a model capable of generalising (we don't want an overfitted model).
Tomorrow we will see how can we check if the parameters we chose are the most appropriate ones or not.
How can we know whether q = 6 is better than q = 3 or not?
Enough for today... Tomorrow I will show you how to know if our initial parameters are the optimal ones.
Follow me on @daansan_ml if you don't want to miss out,
Support my content with a Retweet and Like on the first tweet of this thread.
this twitter thread.