Time-Aware Models Time series data is a sequence of data points

collected at usually constant time intervals.

For example

temperature

humidity

or precipitation measurements recorded every day

or every

day by the weather station, every week, or every hour, or every day by the weather station,

electricity consumption

hourly or daily electricity usage for a household or business

medical

data

for example

patient vital signs

blood pressure

and heart rate recorded at irregular

time intervals during a hospital stay

or maintenance records

equipment maintenance

logs where entries are made only when maintenance is performed.

Time series analysis helps in understanding trends

seasonality

and patterns over time.

Time series forecasting involves predicting the future values based on historical data.

For example, we can predict temperature in a location by considering historical data

at the same location

univariate

or surrounding areas

multivariate.

Some successful forecasting architectures are recurrent neural networks, long short-term

memory

gated recurrent units

and temporal convolutional networks.

Transformers are another successful forecasting architecture.

Traditional time series models often rely on certain assumptions about the data.

These assumptions simplify the modeling process but may not always hold true in real-world

scenarios.

Two major assumptions come to mind.

Constant sampling interval, i.e.

delta t is constant, and prediction interval, same as

the input delta t.

For example

predicting the melting of ice in the Arctic Circle through satellite imagery

for every day for the next three months.

So the assumption is that satellite images are recorded at a high quality every day.

But the reality is that there is a quality drop-off from the satellite imagery and is

recorded only once a week.

The problem is evident.

Data is available every week, but we want to make predictions on a daily basis and sometimes

the data is not even recorded properly.