Datasets:
File size: 15,131 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 | """Step 8 of the ARIMA experiments: rolling-origin backtest on pre-2019 data only.
Origins 1990, 1994, 1998, 2002, 2006, 2011 with a 7-year horizon (backtest.ORIGINS / backtest.H).
No random K-fold: every model sees only data up to its origin.
* Each model's predictions for every (metric, origin) are computed once and cached in
results/arima_backtest/preds/<model>.parquet (delete the folder, or pass --force, to recompute).
* The official MAE and skill S_m come from backtest.run() (unchanged), fed the cached predictions.
* On the same rows (valid years, cities with a 30-year baseline) we add RMSE, MAPE, mean bias
(forecast - actual) and MAE by horizon (1-7 years).
* Composite Score per origin = 100 x max(0, mean of the four S_m), averaged over origins,
as in 02_backtest.py; computed for models that cover all four metrics.
* Reference models: last year repeated, 30-year average, 30-year linear trend, and our current
model (final_model.forecast).
Outputs: results/arima_backtest/{metrics_by_origin.csv, summary.csv, score_by_origin.csv,
mae_by_horizon.csv, fallbacks.csv, *.png} and results/arima_orders.csv.
Run: python code/arima_backtest.py [--force] [--only name1,name2]
"""
import argparse
import time
import warnings
warnings.filterwarnings("ignore")
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import backtest
import final_model
import models
import models_arima as ma
from arima_prep import EDA_CITIES
from backtest import COORDS, H, LV, ORIGINS
from common import METRICS, ROOT, baseline30
OUT = ROOT / "results" / "arima_backtest"
PREDS = OUT / "preds"
ALL = list(METRICS)
def last_year(lv, o, yrs, coords, metric): # identical to 02_backtest.last_year
W = models.wide(lv, metric, o)
last = W.ffill(axis=1).iloc[:, -1]
return pd.concat([pd.DataFrame({"ID": W.index, "metric": metric, "Year": y, "pred": last.values}) for y in yrs])
def P(fn, metrics=ALL, **kw):
return dict(fn=fn, kw=kw, metrics=metrics)
MODELS = {
# reference models
"last_year": P(last_year),
"avg_30yr": P(models.baseline30_model),
"linear_trend": P(models.linear30_model),
"ours": P(final_model.forecast),
# fixed orders proposed by ACF/PACF (Step 5), 50-year window
"arima_000_ct": P(ma.arima_fixed, order=(0, 0, 0), trend="ct"),
"arima_100_ct": P(ma.arima_fixed, order=(1, 0, 0), trend="ct"),
"arima_101_ct": P(ma.arima_fixed, order=(1, 0, 1), trend="ct"),
"arima_100_c": P(ma.arima_fixed, order=(1, 0, 0), trend="c"),
"arima_010_drift": P(ma.arima_fixed, order=(0, 1, 0), trend="c"),
"arima_011_drift": P(ma.arima_fixed, order=(0, 1, 1), trend="c"),
"arima_111_drift": P(ma.arima_fixed, order=(1, 1, 1), trend="c"),
# auto-ARIMA (pmdarima, stepwise AICc, p,q <= 3, d <= 1)
"auto_kpss": P(ma.auto_arima_model, test="kpss", trend0="ct"),
"auto_kpss_c": P(ma.auto_arima_model, test="kpss", trend0="c"),
"auto_adf": P(ma.auto_arima_model, test="adf", trend0="ct"),
"auto_d0_ct": P(ma.auto_arima_d0_trend, trend0="ct"),
"auto_d0_c": P(ma.auto_arima_d0_trend, trend0="c"),
# ARMA around the pooled (all-city, 50-yr) trend rate
"pooled_auto": P(ma.pooled_trend_arima),
"pooled_000": P(ma.pooled_trend_arima, order=(0, 0, 0)),
# training-window length (named by the start year at origin 2018)
"auto_kpss@1949": P(ma.auto_arima_model, test="kpss", trend0="ct", train_start=1949),
"auto_kpss@1979": P(ma.auto_arima_model, test="kpss", trend0="ct", train_start=1979),
"auto_d0_ct@1949": P(ma.auto_arima_d0_trend, trend0="ct", train_start=1949),
"auto_d0_ct@1979": P(ma.auto_arima_d0_trend, trend0="ct", train_start=1979),
"pooled_auto@1949": P(ma.pooled_trend_arima, train_start=1949),
"pooled_auto@1979": P(ma.pooled_trend_arima, train_start=1979),
# CDD from the ARIMA temperature forecast (pooled_auto) + shifted daily weather
"cdd_hybrid_N10": P(ma.cdd_hybrid, metrics=["cdd"], N=10),
"cdd_hybrid_N20": P(ma.cdd_hybrid, metrics=["cdd"], N=20),
}
REFERENCE = ["last_year", "avg_30yr", "linear_trend", "ours"]
# fixed colour per model (validated categorical palette, fixed order); 30-yr average = neutral grey
COLORS = {"ours": "#2a78d6", "arima_011_drift": "#eb6834", "arima_100_ct": "#1baf7a",
"linear_trend": "#eda100", "auto_kpss": "#e87ba4", "auto_d0_ct": "#008300",
"pooled_auto": "#6250d6", "cdd_hybrid_N10": "#e34948", "avg_30yr": "#77776f"}
PLOT_MODELS = list(COLORS)
def years_for(o):
return list(range(o + 1, min(o + H, 2018) + 1))
def compute_preds(name, spec, force=False):
f = PREDS / f"{name.replace('@', '_at_')}.parquet"
if f.exists() and not force:
return pd.read_parquet(f), None
parts, logs = [], []
for m in spec["metrics"]:
for o in ORIGINS:
p = spec["fn"](LV, o, years_for(o), COORDS, m, **spec["kw"])
lg = p.attrs.get("log")
if lg is not None and len(lg):
logs.append(lg.assign(model=name, metric=m, origin=o))
q = p[["ID", "metric", "Year", "pred"]].assign(origin=o)
q.attrs = {} # the log travels separately; pandas compares attrs on concat
parts.append(q)
df = pd.concat(parts, ignore_index=True)
df.to_parquet(f)
return df, (pd.concat(logs, ignore_index=True) if logs else None)
def official(name, preds, metric):
"""MAE and S_m through the unchanged backtest.run(), fed the cached predictions."""
def cached(lv, o, yrs, coords, m):
return preds[(preds.origin == o) & (preds.metric == m) & preds.Year.isin(yrs)][["ID", "metric", "Year", "pred"]]
return backtest.run(cached, metric)
def detail(preds, metric):
"""Per-row errors on exactly the rows common.score() uses."""
rows = []
for o in ORIGINS:
yrs = years_for(o)
lvm = LV[LV.metric == metric]
truth = lvm[lvm.Year.isin(yrs) & lvm.valid][["ID", "metric", "Year", "value"]]
b = baseline30(lvm, o).rename("base").reset_index()
df = truth.merge(b, on=["ID", "metric"]).dropna(subset=["base"])
df = df.merge(preds[(preds.origin == o) & (preds.metric == metric)].drop(columns="origin"),
on=["ID", "metric", "Year"], how="left")
rows.append(df.assign(origin=o, h=df.Year - o, e=df.pred - df.value, eb=df.base - df.value))
return pd.concat(rows, ignore_index=True)
def stats(d):
ae = d.e.abs()
nz = d.value.abs() > 1e-9
return pd.Series(dict(n=len(d), MAE=ae.mean(), RMSE=np.sqrt((d.e ** 2).mean()),
MAPE=100 * (ae[nz] / d.value.abs()[nz]).mean(), bias=d.e.mean(),
MAE_base=d.eb.abs().mean(), S=1 - ae.mean() / d.eb.abs().mean()))
# --------------------------------------------------------------------------- plots
def style(ax):
ax.grid(axis="y", color="#e4e4e0", lw=0.8); ax.set_axisbelow(True)
for s in ("top", "right"):
ax.spines[s].set_visible(False)
ax.spines["left"].set_color("#b8b8b0"); ax.spines["bottom"].set_color("#b8b8b0")
ax.tick_params(colors="#55554f", labelsize=8)
def plot_bars(summary):
fig, axes = plt.subplots(1, 4, figsize=(18, 0.32 * summary.model.nunique() + 2), sharey=True)
order = summary[summary.metric == "annual_avg_temp"].sort_values("S").model.tolist()
order += [m for m in summary.model.unique() if m not in order]
for ax, m in zip(axes, ALL):
s = summary[summary.metric == m].set_index("model").reindex(order)
col = [COLORS.get(k, "#9d9d95") for k in s.index]
ax.barh(range(len(s)), s.S.values, color=col, height=0.7, edgecolor="white", linewidth=1)
ax.axvline(0, color="#55554f", lw=0.8)
ref = s.S.get("ours", np.nan)
if np.isfinite(ref):
ax.axvline(ref, color=COLORS["ours"], lw=1, ls="--")
ax.set_yticks(range(len(s))); ax.set_yticklabels(s.index, fontsize=8)
ax.set_title(f"{m}: mean S_m over 6 origins", fontsize=10); style(ax)
ax.set_xlim(max(-0.4, np.nanmin(s.S.values) - 0.02), None)
for i, v in enumerate(s.S.values):
if np.isfinite(v):
ax.text(v, i, f" {v:.3f}", va="center", fontsize=6.5, color="#2b2b28")
fig.suptitle("Skill vs 30-year average (S_m = 1 - MAE/MAE_30yr); dashed = current model; "
"bars clipped at -0.4", fontsize=11)
fig.tight_layout(); fig.savefig(OUT / "skill_by_model_metric.png", dpi=110); plt.close(fig)
def plot_score_lines(score):
fig, ax = plt.subplots(figsize=(10, 5.5))
for name in PLOT_MODELS:
if name not in score.index:
continue
r = score.loc[name, ORIGINS]
ax.plot(ORIGINS, r.values, marker="o", ms=5, lw=2 if name == "ours" else 1.5,
color=COLORS[name], ls="--" if name == "avg_30yr" else "-", label=f"{name} (mean {score.loc[name, 'mean']:.1f})")
ax.set_xticks(ORIGINS); ax.set_xlabel("forecast origin (forecasts origin+1 .. origin+7)")
ax.set_ylabel("composite Score (0-100)"); style(ax)
ax.legend(fontsize=8, frameon=False, loc="upper left")
ax.set_title("Composite Score by origin", fontsize=11)
fig.tight_layout(); fig.savefig(OUT / "score_by_origin.png", dpi=110); plt.close(fig)
def plot_horizon(hz):
fig, axes = plt.subplots(1, 4, figsize=(18, 4.5))
for ax, m in zip(axes, ALL):
for name in PLOT_MODELS:
s = hz[(hz.model == name) & (hz.metric == m)]
if len(s):
ax.plot(s.h, s.MAE, marker="o", ms=4, lw=1.5, color=COLORS[name],
ls="--" if name == "avg_30yr" else "-", label=name)
ax.set_title(m, fontsize=10); ax.set_xlabel("horizon (years ahead)"); ax.set_xticks(range(1, 8)); style(ax)
axes[0].set_ylabel("MAE (pooled over origins)")
h, l = axes[3].get_legend_handles_labels()
fig.legend(h, l, loc="lower center", ncol=len(l), fontsize=8, frameon=False)
fig.tight_layout(rect=(0, 0.07, 1, 1)); fig.savefig(OUT / "mae_by_horizon.png", dpi=110); plt.close(fig)
def plot_cities(allpreds, cities=("USW00094823", "USW00023183", "USW00024233")):
for cid in cities:
fig, axes = plt.subplots(4, 1, figsize=(13, 13), sharex=True)
for ax, m in zip(axes, ALL):
t = LV[(LV.ID == cid) & (LV.metric == m) & (LV.Year >= 1970) & LV.valid]
ax.plot(t.Year, t.value, color="#2b2b28", lw=1.2, marker=".", ms=4, label="actual")
for name in PLOT_MODELS:
p = allpreds.get(name)
if p is None:
continue
p = p[(p.ID == cid) & (p.metric == m)]
for i, o in enumerate(ORIGINS):
q = p[p.origin == o].sort_values("Year")
ax.plot(q.Year, q.pred, color=COLORS[name], lw=1.5, ls="--" if name == "avg_30yr" else "-",
label=name if i == 0 else None)
for o in ORIGINS:
ax.axvline(o + 0.5, color="#d0d0c8", lw=0.8)
ax.set_title(f"{EDA_CITIES[cid]} - {m}", fontsize=10); style(ax)
h, l = axes[3].get_legend_handles_labels()
fig.legend(h, l, loc="lower center", ncol=len(l), fontsize=8, frameon=False)
fig.suptitle("Actual series and each model's 7-year forecasts from every backtest origin (grey lines = origins)", fontsize=11)
fig.tight_layout(rect=(0, 0.03, 1, 0.98))
fig.savefig(OUT / f"forecasts_{EDA_CITIES[cid].lower()}.png", dpi=100); plt.close(fig)
# --------------------------------------------------------------------------- main
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--force", action="store_true")
ap.add_argument("--only", default="")
args = ap.parse_args()
PREDS.mkdir(parents=True, exist_ok=True)
names = [n for n in MODELS if not args.only or n in args.only.split(",")]
orders_f = ROOT / "results" / "arima_orders.csv"
old_orders = pd.read_csv(orders_f) if orders_f.exists() else pd.DataFrame(columns=["model"])
allpreds, new_logs, rows, dets = {}, [], [], []
for name in names:
t0 = time.time()
preds, lg = compute_preds(name, MODELS[name], args.force)
allpreds[name] = preds
if lg is not None:
new_logs.append(lg)
for m in MODELS[name]["metrics"]:
off = official(name, preds, m)
d = detail(preds, m)
for o in ORIGINS:
s = stats(d[d.origin == o])
r = off[off.origin == o].iloc[0]
assert abs(s.MAE - r.mae) < 1e-9 and abs(s.S - r.S) < 1e-9, (name, m, o)
rows.append(dict(model=name, metric=m, origin=o, **s))
dets.append(d.assign(model=name)[["model", "metric", "origin", "h", "e", "eb"]])
print(f"{name:18s} {time.time() - t0:6.1f}s", flush=True)
if new_logs:
nl = pd.concat(new_logs, ignore_index=True)
old_orders = old_orders[~old_orders.model.isin(nl.model.unique())]
pd.concat([old_orders, nl], ignore_index=True).to_csv(orders_f, index=False)
orders = pd.read_csv(orders_f) if orders_f.exists() else None
mo = pd.DataFrame(rows)
mo.to_csv(OUT / "metrics_by_origin.csv", index=False)
summary = mo.groupby(["model", "metric"], sort=False)[["MAE", "RMSE", "MAPE", "bias", "S"]].mean().reset_index()
s2011 = mo[mo.origin == 2011].set_index(["model", "metric"]).S.rename("S_2011")
summary = summary.join(s2011, on=["model", "metric"])
summary.to_csv(OUT / "summary.csv", index=False)
full = [n for n in names if set(MODELS[n]["metrics"]) == set(ALL)]
score = (mo[mo.model.isin(full)].groupby(["model", "origin"]).S.mean().clip(lower=0).mul(100).unstack())
score["mean"] = score[ORIGINS].mean(1)
score = score.sort_values("mean", ascending=False)
score.to_csv(OUT / "score_by_origin.csv")
det = pd.concat(dets, ignore_index=True)
hz = det.groupby(["model", "metric", "h"]).apply(lambda d: d.e.abs().mean(), include_groups=False).rename("MAE").reset_index()
hz.to_csv(OUT / "mae_by_horizon.csv", index=False)
if orders is not None and len(orders):
fb = (orders.assign(fallback=orders.status != "ok")
.groupby(["model", "metric"]).agg(series=("ID", "size"), fallback=("fallback", "sum"),
weak=("status", lambda s: (s == "weak").sum()),
failed=("status", lambda s: s.str.startswith("failed").sum())))
fb.to_csv(OUT / "fallbacks.csv")
pd.set_option("display.width", 250); pd.set_option("display.max_rows", 200)
print("\nComposite Score by origin:\n", score.round(1).to_string())
print("\nMean S_m by metric (and origin 2011):")
print(summary.pivot(index="model", columns="metric", values="S")[ALL].round(3).to_string())
print(summary.pivot(index="model", columns="metric", values="S_2011")[ALL].round(3).to_string())
if len(names) == len(MODELS):
plot_bars(summary); plot_score_lines(score); plot_horizon(hz); plot_cities(allpreds)
if __name__ == "__main__":
main()
|