arch/arch-ignored-argument
| Rule | arch-ignored-argument |
|---|---|
| Pack | arch |
| Severity | warning |
| Fix | none, edit by hand |
| Library versions | every version |
arch accepts this argument and never reads it for the model or test you asked for, so it has no effect. Pass the setting that does what you mean, such as mean="AR" for lags.
ar = arch_model(returns, lags=2)with_regressors = arch_model(returns, x=x, vol="garch", p=1, q=1)arch3 = arch_model(returns, vol="arch", q=3)adf = ADF(prices, lags=4, method="bic")paths = res.forecast(horizon=5, method="simulation", random_state=np.random.RandomState(0))ar = arch_model(returns, mean="AR", lags=2)with_regressors = arch_model(returns, x=x, mean="ARX", lags=0, vol="garch", p=1, q=1)arch3 = arch_model(returns, vol="arch", p=3)adf = ADF(prices, method="bic")paths = res.forecast(horizon=5, method="simulation", rng=np.random.default_rng(0).standard_normal)What the rule reports
Section titled “What the rule reports”The rule reports an argument that arch accepts and never reads for the model or the test that the call asks for.
| Argument | Reported when the call has |
|---|---|
lags= on arch_model |
mean "constant", "zero" or "ls", or no mean |
x on arch_model |
mean "constant", "zero", "ar" or "har", or no mean |
o=, q= and power= |
vol="arch" or vol="harch" |
power= |
vol "egarch", "aparch", "fiaparch" or "ngarch" |
o= |
vol="figarch" |
method= and max_lags= on ADF, DFGLS or ZivotAndrews |
a number for lags |
random_state= on forecast |
method="simulation" |
rng= on forecast |
method="bootstrap" |
The rule leaves alone a value that changes nothing in a GARCH model either, such as o=0, q=0
or power=2.0. It also leaves alone lags=0, x=None, and a mean or a vol that is a
variable. A call with three or more positional arguments is left alone, because its mean may
be positional.
How to fix it
Section titled “How to fix it”xpd has no automatic fix. Each message names two rewrites, and the call does not show which one is meant, and the rng form needs a seed that the file does not hold. Make the change by hand.
A reported argument does not change the fit. Each form was fitted with and without the argument
on 1,500 daily S&P 500 returns, on arch 9.0.0 and on 8.0.0, and the table gives whether the
parameters of the two fits were equal. "fiaparch" and "ngarch" are new in 9.0 and were run
there only.
| Call | Compared with | Parameters |
|---|---|---|
arch_model(y, lags=2), and with mean="zero" or "ls" |
the same call without lags |
equal |
arch_model(y, mean="ar", lags=2) |
without lags |
differ |
arch_model(y, x=x), and with mean="zero", "ar" or "har" |
without x |
equal |
arch_model(y, x=x, mean="ls"), and with "arx" or "harx" |
without x |
differ |
vol="arch" or "harch" with o=1, q=2 or power=1.0 |
o=0, q=1, power=2.0 |
equal |
vol="egarch", "aparch", "fiaparch" or "ngarch" with power=1.0 |
power=2.0 |
equal |
vol="figarch" with o=1 |
o=0 |
equal |
vol="garch" with o=1, q=2 or power=1.0 |
o=0, q=1, power=2.0 |
differ |
ADF, DFGLS and ZivotAndrews search for a lag length only when lags is not given. With
lags=4 the statistic of 1,000 log prices was the same for method "aic", "bic" and
"t-stat", and the same for max_lags 2 and 20. Without lags, "t-stat" chose 8 lags where
"aic" chose 0, and the statistic of ADF moved from -1.2674 to -1.0838.
forecast reads random_state only for method="bootstrap" and rng only for
method="simulation". The table gives the variance five steps ahead from two calls with the
same seed.
| Method | Seed | First call | Second call |
|---|---|---|---|
"simulation" |
random_state=np.random.RandomState(0) |
2.586 | 2.598 |
"simulation" |
a seeded rng |
2.571 | 2.571 |
"bootstrap" |
a seeded rng |
2.588 | 2.470 |
"bootstrap" |
random_state=np.random.RandomState(0) |
2.595 | 2.595 |
The page on volatility forecasts in arch's own documentation passes random_state to a
simulation. A call that passes both seeds is left alone.
The public code had 21 reports, and none had wrong advice.
| Reports | Code |
|---|---|
| 12 | x= or a positional x with a constant mean. Seven are copies of one function that fits arch_model(resid, x=exog, mean="Constant", ...) as if the regressors were in the model. |
| 4 | vol="ARCH" with q=3 or q=1, where the order of an ARCH model is p |
| 2 | random_state= on a simulation |
| 2 | lags=1 with mean="constant" |
| 1 | a method beside lags=8 |
Of the 10 reports on arch's own code, one is a test that passes max_lags=2.5 to check that it
raises, and its advice is wrong. The other nine are in tests and in the
documentation page on volatility forecasts.
Message
Section titled “Message”xpd prints this message with the issue. A name that starts with $ stands for code from your file.
arch_model ignores lags when mean is "constant", "zero" or "ls", and "constant" is the default, so this model has no lagged terms. A fit with lags=2 gave the same parameters as one without it on arch 9.0.0. Pass mean="AR" to get the lags, or drop lags.Browse the other rules of the arch pack.