The design of statistical software

Mitchell O’Hara-Wild

Materials are available on GitHub

@github/mitchelloharawild/talk-statistical-software-design

Background images from unsplash.com contributors

This presentation is biased

opinionated! 😇

These opinions are shaped by my experiences.

So, who am I?

I like to…

🛠️ Make things 🛠️

🍄🌱Grow food 🍯🥚

📝 Design systems 🎨

💻 Write code 💻

📈 Utilise data 📊

I’m most experienced in designing R packages.

There are many talented users and developers of statistical software here, please share your opinions at any time!

My data analytics experience in R is just a small part of the statistical software ecosystem.

Excel, Tableau, PowerBI, etc.

R is a tiny part of the statistical software landscape.

I’m most experienced in using R and designing R packages.

I will explore statistical software design with increasing specificity toward this very specific area I’m most familiar with.

So yes, this talk is also biased 🙃

The design of statistical software

⚠️ This talk is not about writing good code! ⚠️
(watch Jenny Bryan’s presentation for this.)

✅ It’s about designing good (coding) interfaces ✅

Designing user interfaces

The inputs and outputs for users of statistical software

🌠 Your turn! 🌠

What should be considered when designing a UI for data?

User interface design varies by…

User

Device

Domain

Language

Developer

Intended users

Are you designing for a technical user or not?

Which tools do you are your intended users familiar with?

User interface design varies by…

User

Device

Domain

Language

Developer

Input devices

Output devices

• display
• 2d printing
• 3d printing
• AR/VR
• sound

User interface design varies by…

User

Device

Domain

Language

Developer

Designing software for the domain

Data exploration

Math and probability

Bioinformatics

AI modelling

Time series

User interface design varies by…

User

Device

Domain

Language

Developer

Design variation by language

Syntax

Paradigm

Community

Case study:

Developed by Facebook for R and Python users.

Case study: Prophet for Python

# load software
import pandas as pd
from prophet import Prophet

# data requires time in 'ds' and response in 'y' columns
df = pd.read_csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

Case study: Prophet for Python

# load software
import pandas as pd
from prophet import Prophet

# data requires time in 'ds' and response in 'y' columns
df = pd.read_csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

# specify and estimate model
m = Prophet()
m.fit(df)

Case study: Prophet for Python

# load software
import pandas as pd
from prophet import Prophet

# data requires time in 'ds' and response in 'y' columns
df = pd.read_csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

# specify and estimate model
m = Prophet()
m.fit(df)

# produce forecasts
future = m.make_future_dataframe(periods=365)
forecast = m.predict(future)

Case study: Prophet for Python

# load software
import pandas as pd
from prophet import Prophet

# data requires time in 'ds' and response in 'y' columns
df = pd.read_csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

# specify and estimate model
m = Prophet()
m.fit(df)

# produce forecasts
future = m.make_future_dataframe(periods=365)
forecast = m.predict(future)

# plot results
fig1 = m.plot(forecast)
fig2 = m.plot_components(forecast)

Case study: Prophet for R

# load software
library(prophet)

# data requires time in 'ds' and response in 'y' columns
df <- read.csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

Case study: Prophet for R

# load software
library(prophet)

# data requires time in 'ds' and response in 'y' columns
df <- read.csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

# specify and estimate model
m <- prophet()
m <- fit.prophet(m, df)

(this is slightly changed from the current documentation)

Case study: Prophet for R

# load software
library(prophet)

# data requires time in 'ds' and response in 'y' columns
df <- read.csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

# specify and estimate model
m <- prophet()
m <- fit.prophet(m, df)

# produce forecasts
future <- make_future_dataframe(m, periods = 365)
forecast <- predict(m, future)

Case study: Prophet for R

# load software
library(prophet)

# data requires time in 'ds' and response in 'y' columns
df <- read.csv('https://raw.githubusercontent.com/facebook/prophet/main/examples/example_wp_log_peyton_manning.csv')

# specify and estimate model
m <- prophet()
m <- fit.prophet(m, df)

# produce forecasts
future <- make_future_dataframe(m, periods = 365)
forecast <- predict(m, future)

# plot results
plot(m, forecast)
prophet_plot_components(m, forecast)

Case study: Prophet

Pros:

✅ Very similar interface, easy to switch languages.

✅ Easy to learn the few functions which produce forecasts.

Cons:

❌ Doesn’t follow some R design themes.

❌ Linear code flow makes non-standard analysis hard.

User interface design varies by…

User

Device

Domain

Language

Developer

Design variations by author / group

base R, tidy R, bioconductor

dplyr, ggplot2, shiny

functions or grammars

Designs from the forecast package

That I re-designed for fable

Case study:

Forecasting Australian gas production

Patterns of good software design

Simple and intuitive

Flexible

Consistent

Accessible

Provides feedback

Simple/intuitive: Software should have a simple and intuitive interface that is easy to navigate and use. Users should be able to easily find the features they need and understand how to use them.

Consistency: The design of the software should be consistent throughout. Common functions should return the same structured output.

Helpful feedback: Software should Provide helpful feedback to the user when they perform an action, such as displaying a message when an error occurs. It should also provide guidance to the user when they need help or are unsure of how to perform a task.

Accessibility: Software should be designed to be accessible to users with disabilities. This can include features like text-to-speech and keyboard shortcuts.

Simple and intuitive

library(forecast)
fit <- ets(gas, lambda = 0, model = "AAA")
forecast(fit, h = 72)

         Point Forecast    Lo 80     Hi 80    Lo 95     Hi 95
Sep 1995       54402.25 51048.79  57976.01 49358.08  59961.91
Oct 1995       51192.44 47378.89  55312.94 45476.50  57626.82
Nov 1995       47063.77 43016.54  51491.80 41016.91  54002.09
Dec 1995       43478.22 39276.32  48129.64 37218.95  50790.14
Jan 1996       41047.88 36667.76  45951.22 34541.57  48779.74
Feb 1996       41756.41 36898.20  47254.28 34559.59  50451.92
Mar 1996       46075.15 40285.14  52697.34 37520.72  56579.93
Apr 1996       47375.09 40992.12  54751.97 37969.05  59111.28
May 1996       56195.68 48125.87  65618.65 44334.23  71230.61
Jun 1996       60656.06 51417.55  71554.51 47111.01  78095.50
Jul 1996       65241.81 54745.43  77750.66 49890.93  85315.98
Aug 1996       62433.85 51860.88  75162.36 47009.24  82919.57
Sep 1996       55436.79 45584.78  67418.06 41099.36  74775.79
Oct 1996       52165.94 42462.98  64086.06 38080.04  71462.24
Nov 1996       47958.76 38644.40  59518.14 34470.03  66725.86
Dec 1996       44305.02 35339.13  55545.64 31352.59  62608.36
Jan 1997       41828.46 33025.16  52978.42 29141.87  60038.03
Feb 1997       42550.47 33252.94  54447.59 29184.23  62038.37
Mar 1997       46951.34 36316.72  60700.09 31700.06  69540.18
Apr 1997       48275.99 36957.36  63061.09 32083.24  72641.40
May 1997       57264.32 43385.17  75583.48 37456.73  87546.42
Jun 1997       61809.52 46342.02  82439.58 39788.85  96017.26
Jul 1997       66482.47 49324.51  89608.98 42114.65 104949.67
Aug 1997       63621.12 46705.24  86663.67 39655.68 102069.78
Sep 1997       56491.00 41032.03  77774.19 34643.11  92117.37
Oct 1997       53157.94 38199.97  73973.02 32069.78  88113.08
Nov 1997       48870.76 34742.83  68743.72 29001.56  82352.51
Dec 1997       45147.54 31749.80  64198.84 26351.44  77350.63
Jan 1998       42623.89 29649.68  61275.41 24466.62  74256.11
Feb 1998       43359.62 29831.91  63021.68 24474.24  76817.78
Mar 1998       47844.18 32555.30  70313.16 26552.62  86208.67
Apr 1998       49194.03 33103.20  73106.31 26840.84  90163.05
May 1998       58353.28 38828.95  87695.01 31297.19 108799.06
Jun 1998       62984.92 41440.73  95729.48 33203.47 119478.46
Jul 1998       67746.73 44070.41 104142.87 35098.80 130762.85
Aug 1998       64830.97 41694.29 100806.48 33006.05 127341.94
Sep 1998       57565.25 36597.91  90545.01 28795.69 115078.29
Oct 1998       54168.82 34041.98  86195.35 26620.92 110223.86
Nov 1998       49800.11 30933.75  80172.98 24041.44 103157.34
Dec 1998       46006.08 28243.68  74939.23 21814.81  97023.98
Jan 1999       43434.44 26351.85  71590.83 20226.75  93270.11
Feb 1999       44184.17 26489.97  73697.35 20205.23  96620.58
Mar 1999       48754.01 28882.19  82298.24 21890.83 108582.13
Apr 1999       50129.52 29341.71  85644.94 22097.85 113720.04
May 1999       59462.95 34385.67 102828.94 25731.04 137415.45
Jun 1999       64182.66 36665.30 112351.85 27260.43 151113.30
Jul 1999       69035.03 38956.57 122337.15 28776.54 165615.27
Aug 1999       66063.82 36822.72 118525.42 27023.21 161506.66
Sep 1999       58659.93 32292.49 106556.90 23543.35 146155.43
Oct 1999       55198.91 30010.03 101530.06 21735.10 140184.30
Nov 1999       50747.13 27245.25  94521.84 19601.87 131378.84
Dec 1999       46880.95 24853.46  88431.31 17761.84 123738.53
Jan 2000       44260.41 23167.79  84556.35 16446.10 119115.39
Feb 2000       45024.39 23268.25  87122.83 16405.98 123564.41
Mar 2000       49681.13 25346.73  97378.04 17750.22 139052.62
Apr 2000       51082.80 25726.92 101428.88 17893.53 145832.19
May 2000       60593.72 30122.54 121888.77 20806.99 176459.88
Jun 2000       65403.19 32090.90 133295.63 22013.66 194314.64
Jul 2000       70347.83 34066.00 145271.45 23206.38 213252.41
Aug 2000       67320.11 32171.49 140870.00 21762.90 208244.20
Sep 2000       59775.44 28188.57 126757.15 18934.81 188705.47
Oct 2000       56248.60 26173.10 120883.82 17456.99 181240.03
Nov 2000       51712.15 23740.96 112638.55 15722.51 170083.99
Dec 2000       47772.46 21637.85 105472.94 14227.57 160407.37
Jan 2001       45102.08 20152.70 100939.24 13156.09 154620.31
Feb 2001       45880.59 20222.50 104093.41 13106.56 160608.79
Mar 2001       50625.89 22009.84 116447.02 14161.69 180979.88
Apr 2001       52054.22 22320.71 121395.82 14257.20 190054.29
May 2001       61745.99 26111.89 146008.86 16556.80 230272.09
Jun 2001       66646.92 27794.36 159809.81 17494.05 253904.13
Jul 2001       71685.59 29479.85 174316.46 18417.84 279013.36
Aug 2001       68600.30 27816.73 169178.78 17249.77 272815.26

Simple and intuitive

library(forecast)
fit <- ets(gas, lambda = 0, model = "AAA")
forecast(fit, h = 72)$mean

          Jan      Feb      Mar      Apr      May      Jun      Jul
1995                                                               
1996 41047.88 41756.41 46075.15 47375.09 56195.68 60656.06 65241.81
1997 41828.46 42550.47 46951.34 48275.99 57264.32 61809.52 66482.47
1998 42623.89 43359.62 47844.18 49194.03 58353.28 62984.92 67746.73
1999 43434.44 44184.17 48754.01 50129.52 59462.95 64182.66 69035.03
2000 44260.41 45024.39 49681.13 51082.80 60593.72 65403.19 70347.83
2001 45102.08 45880.59 50625.89 52054.22 61745.99 66646.92 71685.59
          Aug      Sep      Oct      Nov      Dec
1995          54402.25 51192.44 47063.77 43478.22
1996 62433.85 55436.79 52165.94 47958.76 44305.02
1997 63621.12 56491.00 53157.94 48870.76 45147.54
1998 64830.97 57565.25 54168.82 49800.11 46006.08
1999 66063.82 58659.93 55198.91 50747.13 46880.95
2000 67320.11 59775.44 56248.60 51712.15 47772.46
2001 68600.30                                    

Simple and intuitive

library(forecast)
fit <- ets(gas, lambda = 0, model = "AAA")
forecast(fit, h = 72, biasadj = TRUE)$mean

          Jan      Feb      Mar      Apr      May      Jun      Jul
1995                                                               
1996 41207.00 41950.90 46328.12 47677.14 56606.81 61160.26 65852.91
1997 42539.55 43337.88 47894.20 49325.06 58607.38 63370.38 68286.09
1998 44333.42 45205.53 50003.27 51544.27 61301.07 66345.36 71560.00
1999 46736.43 47704.84 52822.61 54507.49 64893.59 70308.14 75915.41
2000 49906.68 50997.28 56531.09 58399.36 69604.81 75496.99 81609.60
2001 54013.01 55255.26 61319.70 63417.01 75669.48 82166.24 88917.18
          Aug      Sep      Oct      Nov      Dec
1995          54469.30 51285.84 47179.62 43614.95
1996 63088.20 56082.95 52838.55 48639.58 44994.59
1997 65471.53 58249.08 54924.93 50602.89 46851.03
1998 68676.73 61160.40 57727.09 53237.67 49340.20
1999 72935.13 65023.12 61439.85 56723.76 52628.92
2000 78493.91 70057.60 66271.37 61253.24 56895.33
2001 85616.84                                    

Simple and intuitive

library(fable)
gas_tsbl |> 
  model(ets = ETS(log(Gas) ~ error("A") + trend("A") + season("A"))) |> 
  forecast(h = "6 years")

# A fable: 72 x 4 [1M]
# Key:     .model [1]
   .model    Month              Gas  .mean
   <chr>     <mth>           <dist>  <dbl>
 1 ets    1995 Sep t(N(11, 0.0025)) 54469.
 2 ets    1995 Oct t(N(11, 0.0036)) 51286.
 3 ets    1995 Nov t(N(11, 0.0049)) 47180.
 4 ets    1995 Dec t(N(11, 0.0063)) 43615.
 5 ets    1996 Jan t(N(11, 0.0078)) 41207.
 6 ets    1996 Feb t(N(11, 0.0093)) 41951.
 7 ets    1996 Mar  t(N(11, 0.011)) 46328.
 8 ets    1996 Apr  t(N(11, 0.013)) 47677.
 9 ets    1996 May  t(N(11, 0.015)) 56607.
10 ets    1996 Jun  t(N(11, 0.017)) 61160.
# ℹ 62 more rows

Simple and intuitive

library(fable)
gas_tsbl |> 
  model(ets = ETS(log(Gas) ~ error("A") + trend("A") + season("A"))) |> 
  forecast(h = "6 years") |> 
  dplyr::mutate(mean(Gas), median(Gas))

# A fable: 72 x 6 [1M]
# Key:     .model [1]
   .model    Month              Gas  .mean `mean(Gas)` `median(Gas)`
   <chr>     <mth>           <dist>  <dbl>       <dbl>         <dbl>
 1 ets    1995 Sep t(N(11, 0.0025)) 54469.      54469.        54402.
 2 ets    1995 Oct t(N(11, 0.0036)) 51286.      51286.        51192.
 3 ets    1995 Nov t(N(11, 0.0049)) 47180.      47180.        47064.
 4 ets    1995 Dec t(N(11, 0.0063)) 43615.      43615.        43478.
 5 ets    1996 Jan t(N(11, 0.0078)) 41207.      41207.        41048.
 6 ets    1996 Feb t(N(11, 0.0093)) 41951.      41951.        41756.
 7 ets    1996 Mar  t(N(11, 0.011)) 46328.      46328.        46075.
 8 ets    1996 Apr  t(N(11, 0.013)) 47677.      47677.        47375.
 9 ets    1996 May  t(N(11, 0.015)) 56607.      56607.        56196.
10 ets    1996 Jun  t(N(11, 0.017)) 61160.      61160.        60656.
# ℹ 62 more rows

Patterns of good software design

Simple and intuitive

Flexible

Consistent

Accessible

Provides feedback

Flexible

library(forecast)
fit <- ets(gas + 1, lambda = 0, model = "AAA")
forecast(fit, h = 72, biasadj = TRUE)

         Point Forecast    Lo 80     Hi 80    Lo 95     Hi 95
Sep 1995       54475.15 51055.26  57981.19 49364.83  59966.68
Oct 1995       51261.05 47358.13  55283.92 45457.62  57595.26
Nov 1995       47181.20 43022.32  51488.87 41024.49  53996.29
Dec 1995       43633.85 39300.36  48142.88 37245.05  50799.58
Jan 1996       41227.27 36695.92  45962.76 34572.82  48785.31
Feb 1996       41998.48 36954.46  47292.00 34618.93  50482.50
Mar 1996       46378.38 40349.81  52731.21 37590.51  56601.90
Apr 1996       47710.57 41048.10  54759.84 38033.19  59100.68
May 1996       56633.31 48187.98  65604.66 44409.12  71187.09
Jun 1996       61192.19 51495.50  71533.67 47204.99  78035.43
Jul 1996       65883.68 54835.51  77712.95 50001.20  85226.55
Aug 1996       63100.73 51942.46  75094.80 47114.00  82790.87
Sep 1996       56105.94 45675.65  67361.49 41212.17  74657.08
Oct 1996       52827.38 42530.57  63982.46 38173.06  71286.16
Nov 1996       48652.91 38733.68  59441.14 34582.62  66576.04
Dec 1996       45023.42 35442.38  55484.87 31477.60  62473.49
Jan 1997       42567.84 33130.87  52912.69 29269.51  59893.16
Feb 1997       43393.00 33388.47  54403.50 29341.04  61908.16
Mar 1997       47950.96 36471.66  60634.87 31880.28  69367.48
Apr 1997       49362.69 37110.05  62955.33 32265.30  72408.29
May 1997       58636.11 43565.52  75422.20 37675.05  87214.42
Jun 1997       63402.26 46549.63  82249.31 40038.90  95623.88
Jul 1997       68313.82 49556.51  89376.51 42394.46 104475.62
Aug 1997       65477.69 46925.75  86395.28 39925.38 101543.50
Sep 1997       58264.27 41246.66  77531.18 34901.35  91626.91
Oct 1997       54902.63 38387.64  73678.28 32303.29  87555.66
Nov 1997       50604.53 34941.45  68486.31 29240.42  81839.14
Dec 1997       46867.47 31953.35  63966.09 26590.82  76866.02
Jan 1998       44347.88 29850.41  61039.52 24701.50  73762.92
Feb 1998       45245.44 30062.36  62801.36 24736.51  76322.68
Mar 1998       50040.52 32815.47  70043.55 26848.55  85610.27
Apr 1998       51558.25 33365.70  72776.73 27142.74  89462.08
May 1998       61297.74 39140.65  87253.51 31657.53 107878.26
Jun 1998       66339.04 41789.76  95224.23 33604.56 118418.35
Jul 1998       71542.34 44454.52 103556.39 35539.21 129534.37
Aug 1998       68634.57 42061.26 100181.52 33428.86 126051.62
Sep 1998       61129.60 36941.20  89975.24 29186.43 113881.46
Oct 1998       57656.27 34352.66  85573.51 26980.20 108956.84
Nov 1998       53192.63 31243.02  79608.69 24391.38 101971.10
Dec 1998       49311.30 28547.53  74416.00 22153.08  95896.05
Jan 1999       46705.15 26646.50  71070.59 20552.83  92142.18
Feb 1999       47696.61 26813.20  73183.34 20555.57  95462.17
Mar 1999       52803.07 29244.09  81691.60 22281.90 107216.93
Apr 1999       54458.25 29709.29  84951.14 22496.92 112185.93
May 1999       64809.94 34821.84 101936.09 26204.92 135455.55
Jun 1999       70210.37 37147.05 111342.71 27780.46 148883.50
Jul 1999       75793.71 39482.13 121188.38 29341.58 163071.50
Aug 1999       72787.00 37324.75 117338.84 27563.30 158894.00
Sep 1999       64894.23 32753.32 105474.47 24033.91 143740.20
Oct 1999       61269.91 30432.28 100400.05 22188.22 137703.81
Nov 1999       56584.92 27653.94  93481.38 20033.08 129043.01
Dec 1999       52510.53 25246.60  87458.24 18171.05 121513.25
Jan 2000       49787.17 23545.35  83597.58 16836.53 116908.55
Feb 2000       50897.30 23672.54  86155.86 16816.92 121278.32
Mar 2000       56405.65 25796.81  96253.88 18205.64 136388.70
Apr 2000       58235.09 26185.01 100179.20 18357.55 142894.51
May 2000       69377.97 30665.16 120310.39 21355.69 172756.69
Jun 2000       75238.69 32685.26 131523.35 22610.52 190127.22
Jul 2000       81308.15 34710.69 143273.91 23850.43 208513.46
Aug 2000       78165.72 32786.53 138839.19 22376.24 203432.53
Sep 2000       69763.91 28746.88 124905.25 19486.11 184266.52
Oct 2000       65937.79 26687.50 118995.18 17966.74 176753.46
Nov 2000       60960.78 24230.89 110887.24 16201.02 165847.38
Dec 2000       56631.60 22103.21 103828.64 14676.56 156368.14
Jan 2001       53751.71 20596.75  99327.38 13581.51 150632.85
Feb 2001       55008.73 20690.96 102451.37 13548.61 156460.10
Mar 2001       61026.84 22529.18 114553.33 14649.03 176174.92
Apr 2001       63073.04 22849.50 119322.51 14752.80 184809.65
May 2001       75221.22 26737.15 143417.47 17140.82 223710.10
Jun 2001       81661.75 28475.34 156911.48 18125.45 246510.12
Jul 2001       88342.35 30215.41 171068.74 19095.75 270683.91
Aug 2001       85017.27 28517.42 165907.52 17893.34 264414.17

Flexible

library(fable)
gas_tsbl |> 
  model(ets = ETS(log(Gas + 1) ~ error("A") + trend("A") + season("A"))) |> 
  forecast(h = "6 years")

# A fable: 72 x 4 [1M]
# Key:     .model [1]
   .model    Month              Gas  .mean
   <chr>     <mth>           <dist>  <dbl>
 1 ets    1995 Sep t(N(11, 0.0025)) 54474.
 2 ets    1995 Oct t(N(11, 0.0036)) 51260.
 3 ets    1995 Nov t(N(11, 0.0049)) 47180.
 4 ets    1995 Dec t(N(11, 0.0063)) 43633.
 5 ets    1996 Jan t(N(11, 0.0077)) 41226.
 6 ets    1996 Feb t(N(11, 0.0093)) 41997.
 7 ets    1996 Mar  t(N(11, 0.011)) 46377.
 8 ets    1996 Apr  t(N(11, 0.013)) 47710.
 9 ets    1996 May  t(N(11, 0.014)) 56632.
10 ets    1996 Jun  t(N(11, 0.016)) 61191.
# ℹ 62 more rows

Patterns of good software design

Simple and intuitive

Flexible

Consistent

Accessible

Provides feedback

Consistent

library(forecast)
fit_ets <- ets(gas + 1, lambda = 0, model = "AAA")
fc_ets <- forecast(fit_ets, h = 72, biasadj = TRUE)
fc_ets

         Point Forecast    Lo 80     Hi 80    Lo 95     Hi 95
Sep 1995       54475.15 51055.26  57981.19 49364.83  59966.68
Oct 1995       51261.05 47358.13  55283.92 45457.62  57595.26
Nov 1995       47181.20 43022.32  51488.87 41024.49  53996.29
Dec 1995       43633.85 39300.36  48142.88 37245.05  50799.58
Jan 1996       41227.27 36695.92  45962.76 34572.82  48785.31
Feb 1996       41998.48 36954.46  47292.00 34618.93  50482.50
Mar 1996       46378.38 40349.81  52731.21 37590.51  56601.90
Apr 1996       47710.57 41048.10  54759.84 38033.19  59100.68
May 1996       56633.31 48187.98  65604.66 44409.12  71187.09
Jun 1996       61192.19 51495.50  71533.67 47204.99  78035.43
Jul 1996       65883.68 54835.51  77712.95 50001.20  85226.55
Aug 1996       63100.73 51942.46  75094.80 47114.00  82790.87
Sep 1996       56105.94 45675.65  67361.49 41212.17  74657.08
Oct 1996       52827.38 42530.57  63982.46 38173.06  71286.16
Nov 1996       48652.91 38733.68  59441.14 34582.62  66576.04
Dec 1996       45023.42 35442.38  55484.87 31477.60  62473.49
Jan 1997       42567.84 33130.87  52912.69 29269.51  59893.16
Feb 1997       43393.00 33388.47  54403.50 29341.04  61908.16
Mar 1997       47950.96 36471.66  60634.87 31880.28  69367.48
Apr 1997       49362.69 37110.05  62955.33 32265.30  72408.29
May 1997       58636.11 43565.52  75422.20 37675.05  87214.42
Jun 1997       63402.26 46549.63  82249.31 40038.90  95623.88
Jul 1997       68313.82 49556.51  89376.51 42394.46 104475.62
Aug 1997       65477.69 46925.75  86395.28 39925.38 101543.50
Sep 1997       58264.27 41246.66  77531.18 34901.35  91626.91
Oct 1997       54902.63 38387.64  73678.28 32303.29  87555.66
Nov 1997       50604.53 34941.45  68486.31 29240.42  81839.14
Dec 1997       46867.47 31953.35  63966.09 26590.82  76866.02
Jan 1998       44347.88 29850.41  61039.52 24701.50  73762.92
Feb 1998       45245.44 30062.36  62801.36 24736.51  76322.68
Mar 1998       50040.52 32815.47  70043.55 26848.55  85610.27
Apr 1998       51558.25 33365.70  72776.73 27142.74  89462.08
May 1998       61297.74 39140.65  87253.51 31657.53 107878.26
Jun 1998       66339.04 41789.76  95224.23 33604.56 118418.35
Jul 1998       71542.34 44454.52 103556.39 35539.21 129534.37
Aug 1998       68634.57 42061.26 100181.52 33428.86 126051.62
Sep 1998       61129.60 36941.20  89975.24 29186.43 113881.46
Oct 1998       57656.27 34352.66  85573.51 26980.20 108956.84
Nov 1998       53192.63 31243.02  79608.69 24391.38 101971.10
Dec 1998       49311.30 28547.53  74416.00 22153.08  95896.05
Jan 1999       46705.15 26646.50  71070.59 20552.83  92142.18
Feb 1999       47696.61 26813.20  73183.34 20555.57  95462.17
Mar 1999       52803.07 29244.09  81691.60 22281.90 107216.93
Apr 1999       54458.25 29709.29  84951.14 22496.92 112185.93
May 1999       64809.94 34821.84 101936.09 26204.92 135455.55
Jun 1999       70210.37 37147.05 111342.71 27780.46 148883.50
Jul 1999       75793.71 39482.13 121188.38 29341.58 163071.50
Aug 1999       72787.00 37324.75 117338.84 27563.30 158894.00
Sep 1999       64894.23 32753.32 105474.47 24033.91 143740.20
Oct 1999       61269.91 30432.28 100400.05 22188.22 137703.81
Nov 1999       56584.92 27653.94  93481.38 20033.08 129043.01
Dec 1999       52510.53 25246.60  87458.24 18171.05 121513.25
Jan 2000       49787.17 23545.35  83597.58 16836.53 116908.55
Feb 2000       50897.30 23672.54  86155.86 16816.92 121278.32
Mar 2000       56405.65 25796.81  96253.88 18205.64 136388.70
Apr 2000       58235.09 26185.01 100179.20 18357.55 142894.51
May 2000       69377.97 30665.16 120310.39 21355.69 172756.69
Jun 2000       75238.69 32685.26 131523.35 22610.52 190127.22
Jul 2000       81308.15 34710.69 143273.91 23850.43 208513.46
Aug 2000       78165.72 32786.53 138839.19 22376.24 203432.53
Sep 2000       69763.91 28746.88 124905.25 19486.11 184266.52
Oct 2000       65937.79 26687.50 118995.18 17966.74 176753.46
Nov 2000       60960.78 24230.89 110887.24 16201.02 165847.38
Dec 2000       56631.60 22103.21 103828.64 14676.56 156368.14
Jan 2001       53751.71 20596.75  99327.38 13581.51 150632.85
Feb 2001       55008.73 20690.96 102451.37 13548.61 156460.10
Mar 2001       61026.84 22529.18 114553.33 14649.03 176174.92
Apr 2001       63073.04 22849.50 119322.51 14752.80 184809.65
May 2001       75221.22 26737.15 143417.47 17140.82 223710.10
Jun 2001       81661.75 28475.34 156911.48 18125.45 246510.12
Jul 2001       88342.35 30215.41 171068.74 19095.75 270683.91
Aug 2001       85017.27 28517.42 165907.52 17893.34 264414.17

Consistent

library(forecast)
fit_ets <- ets(gas + 1, lambda = 0, model = "AAA")
fc_ets <- forecast(fit_ets, h = 72, biasadj = TRUE)
fit_snaive <- snaive(gas)
fc_snaive <- forecast(fit_snaive, h = 72)

Error in forecast.forecast(fit_snaive, h = 72): Please select a longer horizon when the forecasts are first computed

Consistent

library(forecast)
fit_ets <- ets(gas + 1, lambda = 0, model = "AAA")
fc_ets <- forecast(fit_ets, h = 72, biasadj = TRUE)
fit_snaive <- snaive(gas)
# fc_snaive <- forecast(fit_snaive, h = 72)
fit_snaive

         Point Forecast    Lo 80    Hi 80    Lo 95    Hi 95
Sep 1995          57784 54217.08 61350.92 52328.87 63239.13
Oct 1995          53231 49664.08 56797.92 47775.87 58686.13
Nov 1995          50354 46787.08 53920.92 44898.87 55809.13
Dec 1995          38410 34843.08 41976.92 32954.87 43865.13
Jan 1996          41600 38033.08 45166.92 36144.87 47055.13
Feb 1996          41471 37904.08 45037.92 36015.87 46926.13
Mar 1996          46287 42720.08 49853.92 40831.87 51742.13
Apr 1996          49013 45446.08 52579.92 43557.87 54468.13
May 1996          56624 53057.08 60190.92 51168.87 62079.13
Jun 1996          61739 58172.08 65305.92 56283.87 67194.13
Jul 1996          66600 63033.08 70166.92 61144.87 72055.13
Aug 1996          60054 56487.08 63620.92 54598.87 65509.13
Sep 1996          57784 52739.62 62828.38 50069.29 65498.71
Oct 1996          53231 48186.62 58275.38 45516.29 60945.71
Nov 1996          50354 45309.62 55398.38 42639.29 58068.71
Dec 1996          38410 33365.62 43454.38 30695.29 46124.71
Jan 1997          41600 36555.62 46644.38 33885.29 49314.71
Feb 1997          41471 36426.62 46515.38 33756.29 49185.71
Mar 1997          46287 41242.62 51331.38 38572.29 54001.71
Apr 1997          49013 43968.62 54057.38 41298.29 56727.71
May 1997          56624 51579.62 61668.38 48909.29 64338.71
Jun 1997          61739 56694.62 66783.38 54024.29 69453.71
Jul 1997          66600 61555.62 71644.38 58885.29 74314.71
Aug 1997          60054 55009.62 65098.38 52339.29 67768.71

Consistent

library(fable)
gas_tsbl |> 
  model(
    ets = ETS(log(Gas + 1) ~ error("A") + trend("A") + season("A")),
    snaive = SNAIVE(Gas)
  ) |> 
  forecast(h = "6 years")

# A fable: 144 x 4 [1M]
# Key:     .model [2]
   .model    Month              Gas  .mean
   <chr>     <mth>           <dist>  <dbl>
 1 ets    1995 Sep t(N(11, 0.0025)) 54474.
 2 ets    1995 Oct t(N(11, 0.0036)) 51260.
 3 ets    1995 Nov t(N(11, 0.0049)) 47180.
 4 ets    1995 Dec t(N(11, 0.0063)) 43633.
 5 ets    1996 Jan t(N(11, 0.0077)) 41226.
 6 ets    1996 Feb t(N(11, 0.0093)) 41997.
 7 ets    1996 Mar  t(N(11, 0.011)) 46377.
 8 ets    1996 Apr  t(N(11, 0.013)) 47710.
 9 ets    1996 May  t(N(11, 0.014)) 56632.
10 ets    1996 Jun  t(N(11, 0.016)) 61191.
# ℹ 134 more rows

Consistent

Consistency within your own software is a must.

It’s even better to be consistent with your community.

Patterns of good software design

Simple and intuitive

Flexible

Consistent

Accessible

Provides feedback

Accessible

Free and open source

Readily available documentation

Accessible to people of all abilities

Internationalised

Patterns of good software design

Simple and intuitive

Flexible

Consistent

Accessible

Provides feedback

Provide helpful feedback

mu <- mean(gas)

Provide helpful feedback

fc <- gas_tsbl |> 
  model(ets = ETS(log(Gas + 1) ~ error("A") + trend("A") + season("A"))) |> 
  forecast(h = "6 years") |> 
  autoplot(gas_tsbl)

Provide helpful feedback

gas_tsbl |> 
  model(ets = ETS(log(Gas + 1) ~ error("A") + trend("A") + season("A"))) |> 
  forecast(h = "6 years") |> 
  autoplot(gas_tsbl) |> 
  geom_point(colour = "blue")

Error in `geom_point()`:
! `mapping` must be created by `aes()`
ℹ Did you use `%>%` or `|>` instead of `+`?

Some tips for designing good software

Your software design should bee:

Simple

Flexible

Consistent

Accessible

Helpful

Thanks for your time!