cyclisty: like seasonality, but can't be predicted, because the magnitude and the
period changes over time.
And we call this classical decomposition and we call it naive, because it assumes
that the seasonal component is constant over time and we'll get a method that
doesn't assume that in a little bit.
Usually, so I think I have it somewhere here i'll usually start calling a trend,
but it refers to the trend cycle, so we can combine these two components which are
basically just referring to long term changes in the time series that we can't
predict with unknown interval.
Model:
1. Average: avaerage of all observe values
2. Naive: last value of series
3. Sesonally adjusted naive
remove seasonal and then naive
4. seasonal naive:last jan will be future jan, last feb will be future feb
5. Drift:last value + change in slop
6. exponential models:
use all values but weight more on current data
exponentially decaying the weight that we're applying to each previous forecast.
alpha kind of dictates the decay rate rate. the high alpha means the most recent
observations weighted very heavy and then decay decreases very
[Link], if alpha is very small, like 0.1 then it doesn't wait recent
observation is very highly and the decay is very slow.
end or seasonality with SPS it works better on data that don't have A clear trend
or clear seasonal pattern.
a. how do we choose that initial level, well, we have three options we can just set
it to some constant value, like the first observation of the series.
b. what this heuristic is is we fit we use ordinary least squares to fit in line on
the first 10 observations and then we said, oh no, it is the intercept of that
line.
c. Or we can learn it so we can just learn the initial level as a model parameter
and we'll do that a bit later.
All I want you to understand is that simple exponential smoothing is kind of this
in between between the naive and the average forecast. we're using an exponentially
weight at some of our past values to make our forecast so it's just a little bit
smaller.
I mentioned that that's that initial level in an exponential model because doesn't
really matter unless your time series is very short and when I say very short, I
really mean very short it's like 20 observations, would be a short time series
anything more than that.
No seasanality
between naive and average
7. Holt method:drift but based on an exponentially weighted. it now includes a
trend component.
it's tracking over time and it weights more recent trends more heavily than past
trends and then weight is now controlled by beta, which is exactly the same as
alpha that we just saw previously.
8. holt-winter's method:
holds method to now be able to also model seasonality.
it's it's it's kind of like the seasonal naive method but rather than just using
last value we using a
weighted sum of all the last observe values for January or February
It has aaditive and multiple seasonality.
additive dance trend is is basically we damped the trend over time so rather than
just a straight line going forever it kind of flattens out over time, based on a
damping parameter that we can specify.
a. simple exponential= no trend, no seasonality
b. holt method= additive trend but no seasonality.
c. holter winter method-additive / trend and mul seasonality
9. ETS model- can generate distributions and include an error term called ETS
models.
10. selecting the model:
a. Insample: AIC, BIC, MEAN SQUARE ERROR
b. resuduals. no temporal dependecies.
we fit the model to our data we're seeing how well it fits the data that we have
and we're kind of looking at the residuals.
c. OUT SAMPLE: how will our models actually predict unseen data right, how will
they forecast and all these methods.
MAE
RMSE
MAPE
MASE- R SUARED
how will our models actually predict unseen data right, how will they forecast and
all these methods.
the main absolute percentage error and made allows you to compare percentages so
kind of like relative areas which is sometimes very useful for time series which
changing their values at a time.
HIGH ALPHA is close to naive model
heuristic method basically fits a line to the first 10 observations and then set
this value is the intercept of that line.
Arima models:
arena models don't explicitly model, the trend and seasonality and try to encode
memory like that. What they model is the auto correlation in the data Center Member
and lecture one we looked at putting those kilograms and looking at lag features
and that kind of thing.
they coorelate how recent value is correlated with past value. aymond models do is
they're not explicitly modeling trend or seasonality like exponential smoothing
models can. What arena models do is they try to directly model, the temple
dependence, or the auto correlation in the data
- Autoregressive Model(AR): An auto regressive model is just a regression of a
variable against itself right so it's just using as features lagged versions of
itself.
multiplied by some coefficients right so just a very simple linear model with the
explanatory variables as lagged features lag versions of the response.
we usually constrain the parameter the parameters alpha in an order regressive
model to stationary [Link] data is data that doesn't change with time, so
its properties don't change with time, I should say so, like the main or the
variance.
- Moving average Model: What a moving average model is is the rather than using
past values of the response variable in a regression and moving average model uses
positive errors in a regression model.
he model form is exactly the same except rather than being this this linear model
this the sum of coefficient times past values it's a sum of coefficient times past
errors.
Now, unlike whatever aggressive models moving average models a stationary for any
values of the parameters, so we don't have to enforce any constraints on the values
of the parameters can take.
-ARMA: COMBINATION OF THE FIRST TWO model- past values as well as past errors.
stationary and what stationary really is is what it's a time series whose
properties don't depend on time. the time series is roughly horizontal right
there's no trend, it has a fairly constant variance and it doesn't show any
predictable patterns.
- Diferences: different thing and all different thing is is just taking the
difference between consecutive observations so you're essentially going to model,
the change in values rather than the values themselves, so let me show you an
example.
- Seasonal differences: rather than difference in consecutive observations we can
difference, the value of consecutive seasons and i'm going to talk more about that
shortly.
- ARIMA: P,D,Q
-SARIMA: sEASONAL STUFF