Datasets:
File size: 7,210 Bytes
b47743b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 | """Step 7 of the ARIMA experiments: residual diagnostics of the candidate models.
Every candidate is fitted at origin 2018 on the 50-year window 1969-2018 to all 840 series
(weak series skipped, as in the models). For each fit:
* Ljung-Box test on the residuals at lags 5 and 10 (model_df = p + q when smaller than the lag;
the first d residuals, which come from the diffuse start of a differenced model, are dropped);
* AICc. AICc is only comparable between models fitted to the same data: the d=0 models on y,
the d=1 models on the differenced series, the pooled-trend models on the de-trended series.
Figures for the 6 EDA cities (residual line, histogram, ACF) for auto_kpss and pooled_auto.
Outputs: results/arima_diagnostics.csv, results/arima_diagnostics_summary.csv,
results/arima_diagnostics/resid_<city>_<model>.png
Run: python code/arima_diagnostics.py
"""
import warnings
warnings.filterwarnings("ignore")
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from joblib import Parallel, delayed
from statsmodels.stats.diagnostic import acorr_ljungbox
from statsmodels.tsa.stattools import acf
import models
from arima_prep import EDA_CITIES, prepare
from common import CUTOFF_YEAR, METRICS, ROOT, load_annual, long_valid
from models_arima import fit_series
START = 1969
OUT = ROOT / "results" / "arima_diagnostics"
CANDIDATES = {
"arima_000_ct": (dict(kind="fixed", order=(0, 0, 0), trend="ct"), False),
"arima_100_ct": (dict(kind="fixed", order=(1, 0, 0), trend="ct"), False),
"arima_101_ct": (dict(kind="fixed", order=(1, 0, 1), trend="ct"), False),
"arima_100_c": (dict(kind="fixed", order=(1, 0, 0), trend="c"), False),
"arima_010_drift": (dict(kind="fixed", order=(0, 1, 0), trend="c"), False),
"arima_011_drift": (dict(kind="fixed", order=(0, 1, 1), trend="c"), False),
"arima_111_drift": (dict(kind="fixed", order=(1, 1, 1), trend="c"), False),
"auto_kpss": (dict(kind="auto", test="kpss", trend0="ct"), False),
"auto_adf": (dict(kind="auto", test="adf", trend0="ct"), False),
"auto_d0_ct": (dict(kind="auto", d=0, trend0="ct"), False),
"auto_d0_c": (dict(kind="auto", d=0, trend0="c"), False),
"pooled_auto": (dict(kind="auto", d=0, trend0="c"), True),
}
DATA = {"arima_010_drift": "Δy", "arima_011_drift": "Δy", "arima_111_drift": "Δy", "pooled_auto": "y - pooled trend"}
PLOT = ["auto_kpss", "pooled_auto"]
def diag(name, cid, metric, y, spec, keep_resid=False):
try:
m = fit_series(y, spec)
except Exception as e: # noqa: BLE001
return dict(model=name, ID=cid, metric=metric, status=f"failed: {type(e).__name__}")
p, d, q = m.order
r = np.asarray(m.resid())[d:]
out = dict(model=name, ID=cid, metric=metric, status="ok", order=str(m.order), d=d, aicc=m.aicc(),
data="Δy" if (d == 1 and spec["kind"] == "auto") else DATA.get(name, "y"))
for lag in (5, 10):
mdf = p + q if p + q < lag else 0
out[f"lb{lag}_p"] = acorr_ljungbox(r, lags=[lag], model_df=mdf)["lb_pvalue"].iloc[0]
if keep_resid:
out["resid"] = r
return out
def main():
OUT.mkdir(parents=True, exist_ok=True)
lv = long_valid(load_annual())
A = load_annual()
coords = A.groupby("ID")[["Lat", "Lon"]].first()
tasks = []
for metric in METRICS:
S = prepare(lv, metric, CUTOFF_YEAR, START)
slope = models.temp_slope(lv, CUTOFF_YEAR, coords, L=50, metric=metric)
for cid, s in S.items():
if s.weak or len(s.y) == 0:
continue
for name, (spec, pooled) in CANDIDATES.items():
y = s.y - slope[cid] * (s.years - CUTOFF_YEAR) if pooled else s.y
tasks.append((name, cid, metric, y, spec, cid in EDA_CITIES and name in PLOT))
res = Parallel(n_jobs=-1, batch_size=8)(delayed(diag)(*t) for t in tasks)
t = pd.DataFrame(res)
t.drop(columns="resid").to_csv(ROOT / "results" / "arima_diagnostics.csv", index=False)
ok = t[t.status == "ok"]
g = ok.groupby(["model", "metric"])
summ = pd.DataFrame({
"fits": g.size(),
"failed": t[t.status != "ok"].groupby(["model", "metric"]).size(),
"% LB5 p<0.05": g.lb5_p.apply(lambda p: (p < 0.05).mean() * 100),
"% LB10 p<0.05": g.lb10_p.apply(lambda p: (p < 0.05).mean() * 100),
"median AICc": g.aicc.median(),
"data": g.data.agg(lambda s: s.mode().iloc[0]),
}).fillna({"failed": 0})
# AICc wins among d=0 models fitted on y itself (comparable likelihoods)
d0 = ok[ok.model.isin(["arima_000_ct", "arima_100_ct", "arima_101_ct", "arima_100_c", "auto_d0_ct", "auto_d0_c"])]
win = d0.loc[d0.groupby(["ID", "metric"]).aicc.idxmin()].groupby(["metric", "model"]).size()
win = (win / win.groupby(level=0).transform("sum") * 100).rename("% lowest AICc among d=0 on y")
summ = summ.join(win.swaplevel(), how="left")
summ.to_csv(ROOT / "results" / "arima_diagnostics_summary.csv")
pd.set_option("display.width", 250); pd.set_option("display.max_rows", 100)
print(summ.round(1).to_string())
print("\nauto_kpss / auto_adf chosen orders (origin 2018):")
for nm in ("auto_kpss", "auto_adf", "auto_d0_ct", "pooled_auto"):
print(nm, ok[ok.model == nm].groupby("metric").order.agg(lambda s: s.value_counts().head(5).to_dict()).to_dict())
# residual figures for the 6 cities
figs = t[t.get("resid").notna()] if "resid" in t else t.iloc[0:0]
for cid, name in EDA_CITIES.items():
for mdl in PLOT:
fig, ax = plt.subplots(4, 3, figsize=(14, 12))
for i, metric in enumerate(METRICS):
r = figs[(figs.ID == cid) & (figs.model == mdl) & (figs.metric == metric)]
if r.empty:
continue
r = r.iloc[0]
e = r.resid
ax[i, 0].plot(range(START + r.d, START + r.d + len(e)), e, color="#2a78d6", lw=1)
ax[i, 0].axhline(0, color="#55554f", lw=0.8)
ax[i, 0].set_title(f"{metric} {r.order}: residuals", fontsize=9)
ax[i, 1].hist(e, bins=12, color="#2a78d6", edgecolor="white")
ax[i, 1].set_title(f"histogram (LB10 p = {r.lb10_p:.2f})", fontsize=9)
a = acf(e, nlags=10, fft=False)[1:]
band = 1.96 / np.sqrt(len(e))
ax[i, 2].bar(range(1, 11), a, width=0.4, color=np.where(np.abs(a) > band, "#e34948", "#2a78d6"))
ax[i, 2].axhline(band, ls="--", c="#55554f", lw=0.8); ax[i, 2].axhline(-band, ls="--", c="#55554f", lw=0.8)
ax[i, 2].axhline(0, c="#55554f", lw=0.5); ax[i, 2].set_ylim(-0.6, 0.6)
ax[i, 2].set_title("residual ACF (band ±1.96/√N; red = outside)", fontsize=9)
for a_ in ax[i]:
for s in ("top", "right"):
a_.spines[s].set_visible(False)
fig.suptitle(f"{name}: residual diagnostics of {mdl}, fitted on {START}-{CUTOFF_YEAR}", fontsize=11)
fig.tight_layout(); fig.savefig(OUT / f"resid_{name.lower()}_{mdl}.png", dpi=95); plt.close(fig)
if __name__ == "__main__":
main()
|