Seasonality analysis of a coin — seasonality

Do some days of the week, weeks of the year, months or years of the halving cycle bring better returns than others? The returns of the complete periods (see seasonal_returns()) are summed up by group, and a Kruskal-Wallis test asks whether the differences between the groups are larger than chance would give. The test compares the ranks of the returns, so a few extreme months do not decide it, but it takes the returns as independent, which crypto returns are only roughly: calm and wild stretches come in runs.

Usage

seasonality_analysis(history, period = c("day", "week", "month", "year"))

Arguments

history: Price history of one coin from get_history(), with all of it.
period: "day", "week", "month" or "year".

Value

NULL when the history has no complete period with one before it, or a list with:

coin, period: What was analysed.
returns: The returns of seasonal_returns().
groups: A row per group that has returns, in their natural order: group, returns (how many), mean_pct, median_pct, up_pct (the share of the returns above 0), best_pct and worst_pct.
all: The same numbers for all the returns together, as a list.
p_value: The p-value of the Kruskal-Wallis test, NA when fewer than two groups have returns, or no group has two.

Examples

btc <- get_history("Bitcoin (BTC)", as.Date("2013-04-29"))
analysis <- seasonality_analysis(btc, "month")
analysis$groups
#>    group returns   mean_pct median_pct   up_pct   best_pct worst_pct
#> 1    Jan      13 -0.2909088  0.6910592 53.84615  39.835567 -32.08346
#> 2    Feb      13  6.8058464 11.4810969 69.23077  43.716856 -33.81771
#> 3    Mar      13 -0.8480297 -2.1619465 46.15385  30.531110 -32.93329
#> 4    Apr      13  9.2210271  7.5802233 61.53846  34.477899 -17.18056
#> 5    May      14  9.2234063  3.3732947 50.00000  69.628819 -35.35456
#> 6    Jun      14 -1.5659196 -0.5140670 50.00000  26.713348 -37.76883
#> 7    Jul      14  7.7163080  8.1254476 71.42857  23.916334  -8.37172
#> 8    Aug      14  2.3174099 -7.1875290 35.71429  63.576829 -19.17934
#> 9    Sep      14 -2.4603540 -2.4092218 42.85714   7.392924 -19.00922
#> 10   Oct      14 17.5478457 12.9418542 71.42857  53.383459 -12.56590
#> 11   Nov      13 41.8265543  8.7832808 61.53846 453.642183 -36.41158
#> 12   Dec      13  4.0366002 -3.1921079 38.46154  47.773176 -33.23978
analysis$p_value
#> [1] 0.3450331
seasonality_text(analysis)
#> $groups
#> $groups$headline
#> [1] "Best month: November, mean +41.8%; worst: September, −2.5%"
#> 
#> $groups$points
#> [1] "November: mean +41.8%, median +8.8%, up 8 of 13 times (62%)."           "September: mean −2.5%, median −2.4%, up 6 of 14 times (43%)."          
#> [3] "All 162 months together: mean +7.7%, median +2.6%, up 54% of the time."
#> 
#> 
#> $chance
#> $chance$headline
#> [1] "The differences between the months could well be chance (p = 0.35)"
#> 
#> $chance$points
#> [1] "A Kruskal-Wallis test asks whether the returns of the months could all come from the same distribution: p = 0.35, so differences as large as these would turn up in about 35% of histories in which they did not matter."
#> [2] "The returns run from 1 May 2013 to 31 Oct 2026: 162 months, 13 to 14 per group."                                                                                                                                         
#> [3] "With 12 months to pick from, the best one looks good partly by luck, and a pattern of the past need not hold: the crypto market has changed a lot since 2013."                                                           
#> 
#>