Let me show you something you can do with a calculator and a cup of tea. It will change how you read every sales chart you see afterwards.
Here is a consumer electronics retailer’s quarterly sales for three Indian financial years, in ₹ lakh. Quarters run April to June, July to September, October to December, January to March. Year one: 80, 90, 150, 100. Year two: 88, 99, 165, 112. Year three: 97, 108, 180, 124. The October to December quarter, Diwali season, is the tallest every year.
Suppose your manager looks at the 180 and then the next quarter’s 124 and says sales fell by 31%. Did the business get worse? No. 124 is a normal January to March figure. That is the whole reason to decompose a time series: to tell apart three things that get mixed in every chart.
The three ingredients
Trend is the slow, underlying direction: customers are being added, prices drift, the market grows. Seasonality is a pattern that repeats at a fixed interval: festivals, salary weeks, monsoon, exam season, the quarter-end push. Noise is whatever is left over: a rainy Saturday, one big corporate order, a supplier delay.
Step one: smooth out the season
A moving average over exactly one full cycle cancels the seasonal pattern, since every average contains one of each quarter. For quarterly data you average four quarters. Because four is an even number, the window sits between two quarters, so we use a centred version: half of the first and last values, plus the three in between, all divided by four.
For the third quarter of year one (150), the window is half of 80, then 90, 150, 100, and half of 88. That is (40 + 90 + 150 + 100 + 44) / 4 = 106. Doing this along the series gives the trend line: 106.0, 108.1, 111.1, 114.5, 117.1, 119.4, 122.4, 125.8. It climbs gently and without a single festival bump. That is your trend, around 2% a quarter.
Step two: how big is each quarter’s seasonal effect?
Divide each actual figure by its trend value. The October to December quarters give 150 / 106 = 1.415 in year one and 1.409 in year two, and 1.412 averaged out. So the festive quarter runs roughly 41% above trend. Do the same for the others and average across years: April to June about 0.79, July to September about 0.86, January to March about 0.93. The four indices add up to about four, as they should.
These are seasonal indices. They say: a typical April to June quarter lands about 21% below trend, and that is not bad news, it is April.
Step three: use it
To compare quarters fairly, deseasonalise: divide actual by the index. Year three’s three quarters become 97 / 0.79 = about 122.5, 108 / 0.86 = about 125.3, and 180 / 1.41 = about 127.5. After adjustment the business is growing steadily at roughly 2% a quarter, and the festive quarter was a normal one, neither a triumph nor a miracle.
Annual totals tell the same story: 420, 464 and 509 lakh, up about 10.5% and then 9.7%. Comparing the same quarter with the year before (year-on-year) is the quick, honest shortcut, because it holds the season constant. Compare quarter with preceding quarter only after you have adjusted.
Where this goes wrong
Three years is thin. Real analysis wants more cycles to be confident that a seasonal pattern is stable and not a coincidence of two good Octobers. Festival dates also move on the calendar: Diwali can fall in October or November, so a monthly series may see the peak shift. If your data is monthly or daily, calendar quirks (the date of Diwali, number of weekends in a month, trading days) need explicit handling.
Sometimes the seasonal swing grows as the business grows. Here the festive lift was about 41% each year, so the percentage is stable and the multiplicative approach used above fits. If the lift were a constant number of lakh regardless of size, you would subtract instead of divide. Plot the data and check which shape you have.
Finally, a one-off shock, such as a lockdown quarter or a sudden tax change, will distort both the trend and the indices if left in. Mark it, and consider treating it separately. Forecasting methods, from simple exponential smoothing to more elaborate models, are really formal versions of what you just did by hand, so understanding the split makes their output far less mysterious.
Figures and names in the examples are illustrative.
Leave a Reply