tssmooth
Smooths a bumpy series by averaging each point with its neighbours.
What we have
ma, a moving average, withwindow(#l #c #f)
Coming later
exponential,hwintersandshwinters, which are refused- the edge behaviour has not been compared with Stata
You may shorten it to tssmoo.
How you write it
tssmooth ma newvar = var, window(#l #c #f)
Square brackets mean optional. A vertical bar means choose one.
What it does
tssmooth ma trend = sales, window(2 1 2) replaces each month with the average of itself, the two before and the two after — a five-term centred moving average. The month-to-month noise flattens and the underlying movement becomes visible.
The three numbers in window() are how many periods before, the point itself, and how many after.
Watch out. The first and last few points have no full window around them, so they come out missing. That is honest rather than helpful: a smoothed value at the end of a series is a guess.
Try it
Every example below was run after sysuse vira.sample04, clear.
. tsset month
Time variable: month, 1 to 96
Delta: 1 unit
The same thing in pandas and in R
pandas
# tsset: sorted by the time variable; L. and D. count in time, not rows
R
# no R form for `tsset` yet -- see docs/r-equivalents.md
. tssmooth ma smooth = revenue, window(2 1 2)
(4 missing values generated)
The same thing in pandas and in R
pandas
df["smooth"] = df["revenue"].rolling(5, center=True).mean()
R
# no R form for `tssmooth` yet -- see docs/r-equivalents.md
. summarize revenue smooth
Variable | Obs Mean Std. dev. Min Max
-------------+------------------------------------------------------------
revenue | 96 1160.346 350.5008 555.6 2012.2
smooth | 92 1167.145 289.2817 708.44 1845.84
The same thing in pandas and in R
pandas
df[['revenue', 'smooth']].describe()
R
library(dplyr)
df %>% summarise(revenue_n = sum(!is.na(revenue)),
revenue_mean = mean(revenue, na.rm = TRUE),
revenue_sd = sd(revenue, na.rm = TRUE), revenue_min = min(revenue,
na.rm = TRUE), revenue_max = max(revenue, na.rm = TRUE),
smooth_n = sum(!is.na(smooth)), smooth_mean = mean(smooth,
na.rm = TRUE), smooth_sd = sd(smooth, na.rm = TRUE),
smooth_min = min(smooth, na.rm = TRUE), smooth_max = max(smooth,
na.rm = TRUE))
# This R was run in a real R (4.6.1) and gave the same numbers.
The full built-in help (help tssmooth)
tsset timevar [, delta(#)] | tsset panelvar timevar
L.x L2.x F.x D.x D2.x S12.x in any expression
tsline var [var ...]
tssmooth ma newvar = var, window(#l #c #f)
tsdecomp var, period(#) [multiplicative] [generate(stub)]
. tsset month
. generate growth = D.sales / L.sales
. generate yoy = S12.sales / L12.sales
. tsdecomp sales, period(12)
. tsline sales sales_trend
The operators count in units of the time variable, NOT in rows. L.x at
month 5 is x at month 4, and it is missing if there is no month 4 -- even
when the row above is month 3. A plain row shift gets that wrong quietly,
and most real series have a missing month somewhere. In a panel each unit
is lagged on its own.
L. / F. a lag / a lead; L2. is two periods
D. a difference: x - L.x; D2. is a difference of differences
S12. a seasonal difference: x - L12.x
tssmooth ma window(2 1 2) is a 5-term centred average. The edges, where
the window does not fit, are missing
tsdecomp classical decomposition into trend, seasonal and remainder
(created as var_trend, var_seasonal, var_resid). Needs every
period present and two full cycles. Not a Stata command
Not built: date formats (the time variable is a plain number: 1, 2, 3 or
2019, 2020), arima, Holt-Winters, exponential smoothing, dfuller, forecasting.
The decomposition and moving average are checked against R's decompose() and
stats::filter(); the moving average's edge handling has not been compared
with Stata.