{"id":1493,"date":"2026-04-03T18:42:26","date_gmt":"2026-04-03T18:42:26","guid":{"rendered":"https:\/\/inphronesys.com\/?p=1493"},"modified":"2026-04-03T18:42:26","modified_gmt":"2026-04-03T18:42:26","slug":"your-line-chart-is-hiding-8-patterns-how-to-find-them-with-fpp3","status":"publish","type":"post","link":"https:\/\/inphronesys.com\/?p=1493","title":{"rendered":"Your Line Chart Is Hiding 8 Patterns: How to Find Them with fpp3"},"content":{"rendered":"

Four datasets. Same mean. Same variance. Same correlation. Same regression line.<\/p>\n

Completely different shapes.<\/p>\n

This is Anscombe’s Quartet \u2014 a famous statistical demonstration from 1973, and the single best argument for why you should plot your data before you do anything else with it.<\/em> All four datasets produce identical summary statistics: mean of X = 9, mean of Y \u2248 7.5, variance nearly identical, and a correlation coefficient of r \u2248 0.816 across all four. Your ERP’s summary dashboard would show you the same numbers for each one and tell you everything is fine.<\/p>\n

Now look at what happens when you actually plot them:<\/p>\n

\"Anscombe's<\/p>\n

Dataset I is the only one that fits a linear model. Dataset II is a perfect curve \u2014 the linear model is completely wrong. Dataset III has a single outlier dragging the line off course. And Dataset IV? The entire correlation is driven by one extreme point. Remove it and the relationship disappears.<\/p>\n

Same numbers. Four completely different stories. And if you hadn’t plotted them, you’d have treated all four identically \u2014 with a straight line and a confident MAPE.<\/p>\n

This is what happens in supply chain forecasting every single day. Someone pulls 36 months of demand data, calculates an average growth rate, fits a trend, and sends the forecast to the S&OP meeting. Nobody plots it. Nobody checks whether the pattern is linear, curved, seasonal, or driven by a single outlier month when a customer accidentally ordered 10x their normal quantity and then returned half of it.<\/p>\n

"The first thing to do in any data analysis task is to plot the data. Graphs enable many features of the data to be visualised, including patterns, unusual observations, changes over time, and relationships between variables." \u2014 Hyndman & Athanasopoulos, Forecasting: Principles and Practice<\/em><\/p>\n

Hyndman is right. And today, I’m going to show you the visual toolkit that makes plotting supply chain data not just easy, but genuinely fun. By the end of this post, you’ll have eight chart types in your arsenal \u2014 five from fpp3 and three bonus ones that go beyond the textbook \u2014 and you’ll never look at a demand summary table the same way again.<\/p>\n

Your Line Chart Is Lying to You<\/h2>\n

In our first post this month<\/a>, we showed why R crushes Excel for forecasting. Last time<\/a>, you installed R, RStudio, and fpp3, and produced your first real forecast. But we skipped a step. A critical one.<\/p>\n

We went straight from data to model. We didn’t look<\/em> at the data first.<\/p>\n

That’s like a mechanic replacing your engine without popping the hood. Sure, a new engine might fix the problem. But what if the issue was a loose spark plug?<\/p>\n

Every forecasting model makes assumptions about your data. ETS assumes a certain error structure. ARIMA assumes stationarity (or that differencing achieves it). Even Seasonal Naive assumes the seasonal pattern is stable. If those assumptions are wrong \u2014 if your data has a structural break, or an outlier, or a changing seasonal pattern \u2014 the model will happily fit itself to garbage and give you a confident-looking garbage forecast.<\/p>\n

Visualization is how you check those assumptions before<\/em> you bet your inventory on them. And fpp3 gives you a visual toolkit specifically designed for time series data \u2014 not just generic scatter plots, but purpose-built charts that each reveal a different hidden pattern.<\/p>\n

Let’s meet the toolkit. We’ll use Australian quarterly beer production as our primary dataset for the core fpp3 charts \u2014 so you can see how each visualization reveals something new about data you’ve already "seen." Then we’ll explore three creative techniques with different data to show their unique strengths.<\/p>\n

The fpp3 Visual Toolkit<\/h2>\n

1. autoplot() \u2014 The Time Plot<\/h3>\n

The most basic visualization, and the one most teams skip. autoplot()<\/code> plots your time series as a line chart with proper time axis formatting. Two lines of code:<\/p>\n

beer <- aus_production |>\n  filter(year(Quarter) >= 1992) |>\n  select(Quarter, Beer)\n\nbeer |> autoplot(Beer)\n<\/code><\/pre>\n
\"Time<\/p>\n
What this tells a supply chain pro:<\/strong> You immediately see three things \u2014 the overall level (roughly 400\u2013500 megalitres per quarter), a slight downward trend over time, and a strong seasonal pattern that repeats every year. That seasonal pattern is Q4-dominant \u2014 Australians drink more beer in their summer (October\u2013December). You also see that the variability is roughly constant \u2014 the seasonal swings don’t get bigger or smaller, which tells you additive seasonality is probably appropriate.<\/p>\n
This is the chart you should have run before that S&OP meeting. It takes ten seconds and saves you from building a forecast on data you haven’t actually looked at.<\/p>\n
2. gg_season() \u2014 The Seasonal Overlay<\/h3>\n
Now we go deeper. gg_season()<\/code> takes each year of data and overlays them on top of each other, all sharing the same x-axis (Q1 through Q4). This lets you compare the seasonal shape across years instantly:<\/p>\n
beer |> gg_season(Beer, labels = "both")\n<\/code><\/pre>\n
\"Seasonal<\/p>\n
What this tells a supply chain pro:<\/strong> Is the seasonal pattern stable? Getting stronger? Shifting? This chart answers all three at a glance. You can see that Q4 is consistently the peak and Q2 is consistently the trough \u2014 but the peak has been getting slightly lower over the years. The overall "shape" of the season is stable, but the level is declining. If your seasonal factors in Excel haven’t been updated since 2005, this chart tells you they need revision.<\/p>\n
This is how you discover that your December spike has been creeping into November for three years. Or that your Q3 pre-build has shifted from August to July. Or that one year’s pattern was completely anomalous \u2014 and your model is treating it as normal.<\/p>\n
Your ERP vendor charges six figures for a dashboard that can’t do gg_season()<\/code>.<\/p>\n
3. gg_subseries() \u2014 The Subseries Means<\/h3>\n
This one is less well known, and it’s a gem. gg_subseries()<\/code> splits your data into mini-panels \u2014 one per season (quarter, in our case). Within each panel, you see the observations for that quarter over time. And the horizontal blue line shows the mean for that period:<\/p>\n
beer |> gg_subseries(Beer)\n<\/code><\/pre>\n
\"Subseries<\/p>\n
What this tells a supply chain pro:<\/strong> Is each season trending the same way, or are some quarters changing while others hold steady? Look at Q4 \u2014 it’s clearly declining. But Q2 might be more stable. That distinction matters enormously for planning: if your peak season is eroding while your trough stays flat, your overall demand is contracting, and your seasonal factors need to shrink the peak amplitude.<\/p>\n
This is the chart that tells you Q4 production has been declining for 15 years. And makes you ask: did anyone notice? Was this deliberate (market shift, health trends, category decline) or did it slip under the radar while everyone focused on the annual total?<\/p>\n
4. gg_lag() \u2014 The Lag Scatter Plot<\/h3>\n
This chart asks: does last quarter’s production predict this quarter’s production? And if so, which lags matter most?<\/p>\n
beer |> gg_lag(Beer, lags = 1:9)\n<\/code><\/pre>\n
\"Lag<\/p>\n
What this tells a supply chain pro:<\/strong> Each panel plots the current quarter’s value (y-axis) against the value k quarters ago (x-axis). If you see a tight cluster along the diagonal, that lag is a strong predictor. The color coding shows which quarter each point belongs to.<\/p>\n
Look at lag 4 \u2014 that’s the "same quarter last year" comparison. If that panel shows a tight linear relationship, it means last year’s Q4 is a good predictor of this year’s Q4. That’s Seasonal Naive in visual form. If lag 1 also shows a tight relationship, recent momentum matters too. If only lag 4 is tight and lag 1 is scattered, your data is seasonal but not autoregressive \u2014 a critical distinction for model selection.<\/p>\n
This is the chart that tells you whether "last quarter predicts this quarter" or "same quarter last year predicts this quarter." Your model choice depends on the answer. And yes, we just diagnosed a model selection problem by looking at a scatter plot of beer production. Sometimes the best tools are the simplest \u2014 and the ones that pair well with an actual beer.<\/p>\n
5. ACF Plot \u2014 The Autocorrelation Fingerprint<\/h3>\n
The ACF (AutoCorrelation Function) plot is the single most information-dense chart in time series analysis. It summarizes the correlation structure of your entire dataset in one compact graph:<\/p>\n
beer |> ACF(Beer) |> autoplot()\n<\/code><\/pre>\n
\"ACF<\/p>\n
What this tells a supply chain pro:<\/strong> Read it like a diagnostic scan. Three patterns to look for:<\/p>\n