Step by Step
T
Trend — the long-term direction
The general upward, downward, or flat direction a time series moves over a long period, ignoring short-term fluctuations.
Example: a company's revenue steadily climbing over five years despite month-to-month ups and downs.
A
Autocorrelation — correlation with its own past
The degree to which a value in a time series correlates with its own earlier values, which is a defining feature that separates time series data from ordinary independent data.
Example: today's stock price is often correlated with yesterday's price, unlike two unrelated rows in a typical tabular dataset.
S
Seasonality — regular repeating patterns
Predictable, repeating patterns tied to a fixed calendar cycle, such as daily, weekly, or yearly patterns.
Example: retail sales spiking every December due to holiday shopping, repeating predictably every year.
C
Cycle — irregular, multi-year patterns
Longer-term fluctuations that repeat but not on a fixed calendar schedule, unlike seasonality's predictable timing.
Example: economic boom-and-bust cycles that recur over several years but without a fixed, predictable interval.
Applied Walkthrough
1
A student randomly shuffles their time series dataset before splitting it into training and test sets, just like they would for a typical tabular dataset.
2
Ask: does this preserve the chronological relationship that time series data depends on? No — shuffling destroys the time order entirely.
3
Worse, this could let information from the future end up in the training set, letting the model implicitly "see" outcomes it should never have access to when predicting the past.
4
The correct approach: always split chronologically — train on earlier data, test on strictly later data — never shuffle time series before splitting.
Exam Application
Exams heavily test the never-shuffle rule specifically because it's such a common and consequential mistake, and because it's easy to forget when applying otherwise-standard ML practices to time series data. Also expect questions distinguishing classic models (ARIMA) from modern approaches (LSTMs, Temporal CNNs, Transformer-based forecasters).
⚠ Common Trap
The single most important trap: shuffling time series data before splitting into train/test, which is standard practice for ordinary tabular data but a serious error here. This can leak future information into training, producing misleadingly optimistic performance that will not hold up in real-world deployment where you genuinely cannot see the future.
✓ Quick Self-Check
1. What does "trend" refer to in a time series?
The long-term upward, downward, or flat direction of the data over an extended period.
Tap to reveal / hide
2. What is autocorrelation?
The degree to which a value in the series correlates with its own earlier (past) values.
Tap to reveal / hide
3. How is seasonality different from a cycle?
Seasonality follows a fixed, predictable calendar schedule (daily/weekly/yearly); cycles are longer-term and irregular, without a fixed schedule.
Tap to reveal / hide
4. Why should you never shuffle time series data before splitting into train/test?
Shuffling destroys the chronological order and can leak future information into training, producing misleadingly optimistic results.
Tap to reveal / hide
5. Name one classic and one modern approach to time series forecasting.
ARIMA is a classic approach; LSTMs, Temporal CNNs, or Transformer-based forecasters are modern approaches.
Tap to reveal / hide