L7: Time Series Forecating Flashcards
What is the goal of time series forecating?
Predict future state of values based on historical values
What is unique about time series forecasting vs. regression or classification?
A) Time dependency and sequential data rather than individual draws
B) Feature engineering: use lag values (past values) as features and time-based features
C) Different evaluation metrics: mean absolute error, root mean squared error etc.
D) Time-based cross validation or walk-forward validation
E) Interdependence between observations
F) Different data partitioning: test and training “periods” rather than “data”
G) Captures seasonality and trends
All options are true
The longer the horizon, the larger the uncertainty of the forecast, leading to the benefit of forecast updating
TRUE/ FALSE
TRUE
Time series ANALYSIS is descriptive as it describes the patterns in historical data.
Time series FORECASTING is predictive, as it predicts future states of values
What does residual error mean in times series forecasting?
Residual error: difference between forecast value vs. actual value
What does “TREND” in times series mean?
Whether the series has an overall increasing/ decreasing trend
What does “SEASONALITY” in times series mean?
Whether any spikes repat itself at fixed intervals (seasonality) or non-fixed interval (cycle)
What does “RANDOM NOISE” in times series mean?
Any form of observation left after accounting for trend and seasonality
What is global trend vs. local trends?
Local trend: Represents shorter-term fluctuations or patterns within data
Global trend: represents the overall long-term direction of data - captures the broad movement/ pattern
Exponential Smoothing means that observations in the past are weighted depending on how far back they are in time (closer in time observations are assigned higher weights).
TRUE/FALSE
TRUE
The degree of weighting depends on the domain: whether it is fast-changing or fairly constant over time
TRUE/FALSE
TRUE
The degree of decay of weights back in time depends on how important newer data is
Data partitioning in time series forecast: e.g., fixed partitioning implies that you use all past data as training dataset to forecast the future
TRUE/FALSE
TRUE
What are the common data pre-processing considerations in TSF? Select all correct
A) Missing values: some models, e.g., ARIMA cannot handle missing values
B) Unequally spaced series: equal intervals between datapoints is necessary, if not, imputation is required (e.g., setting N.A. to 0)
C) Choice of time span: older data is less informative, but important to have sufficient dataset size
ALL ARE CORRECT:
A) Missing values: some models, e.g., ARIMA cannot handle missing values
B) Unequally spaced series: equal intervals between datapoints is necessary, if not, imputation is required (e.g., setting N.A. to 0)
C) Choice of time span: older data is less informative, but important to have sufficient dataset size
Sliding window is one partitioning technique in TSF. What does this mean?
A) Older data is dropped due to the assumption that newer data is more representative - smaller training set
B) Longer training periods are created as more data is revealed with to be incorporated into the model
A) Older data is dropped due to the assumption that newer data is more representative - smaller training set
Expanding window is one partitioning technique in TSF. What does this mean?
A) Older data is dropped due to the assumption that newer data is more representative - smaller training set
B) Longer training periods are created as more data is revealed with to be incorporated into the model
B) Longer training periods are created as more data is revealed with to be incorporated into the model