Explain like I'm five · AI Theory

What is Time-Series Analysis (AR/MA)?

AR/MA time-series analysis, told as a story about an inspector who guesses tomorrow's harvest from the last few days' numbers and how far recent guesses missed.

Mo Tu We Th Fr Sa Suhow much will come in next week?the harvest count bounces every day — so the next day's count is hard to guess

The daily harvest count bounces around.The harvest count bounces every day — so the next day's count is hard to guess.

yesterday was 24, so today is 24 too — just copying one day backthis forecast misses a lot — one prior day can't capture the ups and downs

Just copying yesterday's number often misses.One prior day can't capture the ups and downs.

AR/MA time-series analysis = blending the last few days and correcting with recent errors to estimate the nextthe green dashed line is the forecast, bars are the real count — they never match exactly

AR/MA time-series analysis estimates the next value from the last few days plus recent errors.The green dashed line is the forecast, bars are the real count — they never match exactly.

Weighted past values (AR)recent days often count more
Recent-error correction (MA)misses feed into the next guess
Check for seasonalitylook for holiday/harvest-season patterns
The distant past often fadesolder values often have less influence
AR PART — WEIGHTED PAST VALUEStoday ≈ yesterday×0.6 + day before×0.3 + …weights come from data, often larger for recent daysMA PART — RECENT-ERROR CORRECTIONhow far off we were feeds into the next guessif errors stay big, revisit the modelcombine AR+MA into ARMA; add seasonality and you get SARIMA

AR is a weighted sum of past values; MA corrects using recent errors.Combine AR and MA for ARMA; add seasonality for SARIMA.

in calm times the forecast tracks wella sudden cold snap makes past patterns uselesstime-series analysis reads ordinary days well — sudden events need separate planning

Time-series forecasts are weak against sudden events.Time-series analysis reads ordinary days well — sudden events need separate planning.

In one breath

AR/MA time-series analysis blends recent past values with weights to estimate what's next (AR), then corrects using how far recent forecasts missed (MA).

AR (AutoRegressive) models the current value as a linear combination of the past p values; MA (Moving Average) models it as a linear combination of past white-noise error terms. Together they form ARMA(p,q); differencing a non-stationary series to make it stationary gives ARIMA(p,d,q); adding seasonality gives SARIMA. The core assumption — that past patterns persist into the future — makes these models fragile against structural breaks like sudden events.

When grown-ups say it

AR (AutoRegressive)
A weighted sum of the last few days. Weights are estimated from data, often larger for recent values.
MA (Moving Average)
A correction from recent forecast errors. Multiplies how far off you were by a weight and feeds it into the next guess.
Stationarity
The assumption that the pattern doesn't shift over time. AR/MA only fits well when the pattern holds steady.
Seasonality
A pattern that repeats, like holidays or harvest season. If present, add a seasonal term as in SARIMA.
Regression Assumptions
The conditions that keep the scale hanging straight. Time series is a cousin of regression and shares similar conditions. → what keeps the scale hanging straight
Unbiased Estimator
A scale dial with no systematic tilt. How a forecast's coefficients are estimated is also checked for bias. → the dial with no systematic tilt
Cross Validation
Splitting into groups and testing in rotation. Time-series forecasts are cross-validated over past windows too. → see it in full
Reinforcement Learning
A sorter that learns from rewards instead of an answer key. A different style of learning from forecasting. → see it in full
Next story
The Certification Board That One Accuracy Number Can't Earn →
Want another term as a picture book? Request →