File size: 14,264 Bytes
c0058c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Reusable pieces for the Protostar climate forecasts.

Everything here works on a "panel": a DataFrame indexed by (city_id, year) with one
column per metric, NaN where the challenge would not count the value. Forecasters are
plain functions `f(panel, meta, origin, horizon_years) -> DataFrame(city_id, year, metric,
predicted_value)` that only look at rows with year <= origin, so the same code runs the
backtests (origin 2004, 2011, ...) and the real forecast (origin 2018).
"""

from __future__ import annotations

from pathlib import Path

import numpy as np
import pandas as pd

ROOT = Path(__file__).resolve().parent.parent
DATA = ROOT / "data"
CUTOFF = 2018  # data rule: no weather after 2018-12-31

METRICS = ["annual_avg_temp", "total_precip_in", "hottest_day_temp", "cdd"]
SOURCE_COLUMN = {
    "annual_avg_temp": "Avg Temp",
    "total_precip_in": "ToT Prcp",
    "hottest_day_temp": "Max Tmax",
    "cdd": "cdd",
}
MAX_MISSING = 10  # challenge rule: a year counts if <= 10 days are missing


# --------------------------------------------------------------------------- data

def load_panel(data_dir: Path = DATA) -> tuple[pd.DataFrame, pd.DataFrame]:
    """Return (panel, meta). panel: (city_id, year) x METRICS with invalid years as NaN,
    plus `segment` (station segment). meta: one row per city (Name, Lat, Lon)."""
    annual = pd.read_parquet(data_dir / "annual_indices.parquet")
    cover = pd.read_parquet(data_dir / "coverage_and_stations.parquet")
    df = annual.merge(cover, on=["ID", "Name", "Year"], how="left")
    if df["Year"].max() > CUTOFF:
        raise RuntimeError(f"Data rule: found weather dated after {CUTOFF}; refusing to continue.")

    days = np.where(pd.to_datetime(df["Year"].astype(str) + "-12-31").dt.is_leap_year, 366, 365)
    temp_ok = (df["Miss Tmax"] <= MAX_MISSING) & (df["Miss Tmin"] <= MAX_MISSING)
    valid = {
        "annual_avg_temp": temp_ok,
        "total_precip_in": df["Miss Prcp"] <= MAX_MISSING,
        "hottest_day_temp": temp_ok,
        "cdd": temp_ok & (days - df["Days Both"] <= MAX_MISSING),
    }
    panel = pd.DataFrame({"city_id": df["ID"], "year": df["Year"], "segment": df["Station Segment"]})
    for m in METRICS:
        v = df[SOURCE_COLUMN[m]].replace([np.inf, -np.inf], np.nan)
        panel[m] = v.where(valid[m])
    # extra series (not scored): annual_avg_temp = (tmax + tmin) / 2
    panel["tmax"] = df["Avg Tmax"].where(temp_ok)
    panel["tmin"] = df["Avg Tmin"].where(temp_ok)
    panel = panel.set_index(["city_id", "year"]).sort_index()
    meta = annual.groupby("ID")[["Name", "Lat", "Lon"]].last()
    meta.index.name = "city_id"
    return panel, meta


def wide(panel: pd.DataFrame, metric: str, origin: int) -> pd.DataFrame:
    """cities x years matrix of one metric, years <= origin only."""
    p = panel[panel.index.get_level_values("year") <= origin]
    return p[metric].unstack("year")


# --------------------------------------------------------------------------- helpers

def _window(w: pd.DataFrame, origin: int, k: int) -> pd.DataFrame:
    cols = [y for y in w.columns if origin - k < y <= origin]
    return w[cols]


def _long(pred: dict[str, pd.DataFrame]) -> pd.DataFrame:
    out = []
    for m, frame in pred.items():
        s = frame.stack()
        s.index.names = ["city_id", "year"]
        out.append(s.rename("predicted_value").reset_index().assign(metric=m))
    return pd.concat(out, ignore_index=True)[["city_id", "year", "metric", "predicted_value"]]


def _broadcast(level: pd.Series, years) -> pd.DataFrame:
    return pd.DataFrame({y: level for y in years})


def segment_offsets(panel: pd.DataFrame, metric: str, origin: int, min_years: int = 5) -> pd.Series:
    """Additive correction putting each (city, year) on the level of the station in use at
    `origin`. For each station change, offset = mean(after) - mean(before) using up to 10
    years each side, with the regional mean of the same years subtracted so the climate
    signal is not mistaken for a station jump."""
    p = panel[panel.index.get_level_values("year") <= origin]
    w = p[metric].unstack("year")
    seg = p["segment"].unstack("year")
    anom = w - w.mean(axis=1).values[:, None]
    national = anom.mean(axis=0)  # simple common signal
    resid = anom - national
    adj = pd.DataFrame(0.0, index=w.index, columns=w.columns)
    for city in w.index:
        s = seg.loc[city].dropna()
        if s.empty:
            continue
        cur = s.iloc[-1]
        changes = s.index[s.diff().fillna(0) != 0]
        r = resid.loc[city]
        for yr in sorted(changes, reverse=True):
            before = r[(r.index < yr) & (r.index >= yr - 10)].dropna()
            after = r[(r.index >= yr) & (r.index < yr + 10)].dropna()
            if len(before) < min_years or len(after) < min_years:
                continue
            jump = after.mean() - before.mean()
            adj.loc[city, adj.columns < yr] += jump
        _ = cur
    return adj


# --------------------------------------------------------------------------- forecasters

def f_mean(panel, meta, origin, years, k=30, metrics=METRICS):
    """Mean of the last k valid years (k=30 is the challenge's baseline)."""
    return _long({m: _broadcast(_window(wide(panel, m, origin), origin, k).mean(axis=1), years) for m in metrics})


def f_last(panel, meta, origin, years, metrics=METRICS):
    out = {}
    for m in metrics:
        w = wide(panel, m, origin)
        out[m] = _broadcast(w.ffill(axis=1).iloc[:, -1], years)
    return _long(out)


def f_city_trend(panel, meta, origin, years, k=30, metrics=METRICS):
    """Per-city OLS line over the last k years, extrapolated (the 'Linear Trend' baseline)."""
    out = {}
    for m in metrics:
        w = _window(wide(panel, m, origin), origin, k)
        x = np.array(w.columns, float)
        preds = {}
        for city, row in w.iterrows():
            ok = row.notna().values
            if ok.sum() < 5:
                preds[city] = [row.mean()] * len(years)
                continue
            b, a = np.polyfit(x[ok], row.values[ok], 1)
            preds[city] = [a + b * y for y in years]
        out[m] = pd.DataFrame(preds, index=list(years)).T
    return _long(out)


def regional_slope(w: pd.DataFrame, meta: pd.DataFrame, origin: int, k: int, radius_km: float) -> pd.Series:
    """Trend (units/yr) per city estimated from the pooled anomalies of all cities within
    `radius_km`. Pooling makes the slope far less noisy than a single-station fit."""
    win = _window(w, origin, k)
    anom = win.sub(win.mean(axis=1), axis=0)
    x = np.array(win.columns, float)
    lat = np.radians(meta.loc[win.index, "Lat"].values)
    lon = np.radians(meta.loc[win.index, "Lon"].values)
    d = 6371 * np.arccos(np.clip(
        np.sin(lat)[:, None] * np.sin(lat)[None, :]
        + np.cos(lat)[:, None] * np.cos(lat)[None, :] * np.cos(lon[:, None] - lon[None, :]), -1, 1))
    slopes = {}
    A = anom.values
    for i, city in enumerate(win.index):
        nb = d[i] <= radius_km
        series = np.nanmean(A[nb], axis=0)
        ok = ~np.isnan(series)
        slopes[city] = np.polyfit(x[ok], series[ok], 1)[0] if ok.sum() >= 5 else 0.0
    return pd.Series(slopes)


def f_level_trend(panel, meta, origin, years, *, level_k=None, trend_k=None, radius_km=None,
                  damp=None, adjust_stations=None, metrics=METRICS, params=None):
    """Level = mean of the last `level_k` years (station-adjusted if asked), plus a regionally
    pooled trend fitted on the last `trend_k` years, applied from the level window's centre
    to each target year and multiplied by `damp` (0 = no trend, 1 = full trend).

    `params` may give per-metric dicts overriding the keyword defaults."""
    defaults = dict(level_k=15, trend_k=40, radius_km=800, damp=1.0, adjust_stations=True)
    out = {}
    for m in metrics:
        cfg = dict(defaults)
        cfg.update({k: v for k, v in dict(level_k=level_k, trend_k=trend_k, radius_km=radius_km,
                                          damp=damp, adjust_stations=adjust_stations).items() if v is not None})
        if params and m in params:
            cfg.update(params[m])
        if cfg.get("trend_start"):  # fixed start year instead of a fixed window length
            cfg["trend_k"] = origin - cfg["trend_start"] + 1
        w = wide(panel, m, origin)
        if cfg["adjust_stations"]:
            w = w + segment_offsets(panel, m, origin)
        lvl_win = _window(w, origin, cfg["level_k"])
        level = lvl_win.mean(axis=1)
        centre = lvl_win.apply(lambda r: np.mean([c for c, v in r.items() if pd.notna(v)]) if r.notna().any() else origin, axis=1)
        slope = regional_slope(w, meta, origin, cfg["trend_k"], cfg["radius_km"]) if cfg["damp"] else 0 * level
        out[m] = pd.DataFrame({y: level + cfg["damp"] * slope * (y - centre) for y in years})
        if m in ("cdd", "total_precip_in"):
            out[m] = out[m].clip(lower=0)
    return _long(out)


def f_cdd_from_temp(panel, meta, origin, years, *, temp_params=None, level_k=15, beta_k=30):
    """CDD driven by the temperature forecast: each city's CDD moves with its annual mean
    temperature at a sensitivity beta (°F-days per °F) fitted on its last `beta_k` years.
    pred = mean CDD (last level_k yrs) + beta * (temp forecast - mean temp (last level_k yrs))."""
    temp = f_level_trend(panel, meta, origin, years, metrics=["annual_avg_temp"],
                         params={"annual_avg_temp": temp_params} if temp_params else None)
    tw = wide(panel, "annual_avg_temp", origin) + segment_offsets(panel, "annual_avg_temp", origin)
    cw = wide(panel, "cdd", origin) + segment_offsets(panel, "cdd", origin)
    t_win, c_win = _window(tw, origin, beta_k), _window(cw, origin, beta_k)
    beta = {}
    for city in t_win.index:
        ok = t_win.loc[city].notna() & c_win.loc[city].notna()
        beta[city] = np.polyfit(t_win.loc[city][ok], c_win.loc[city][ok], 1)[0] if ok.sum() >= 10 else 0.0
    beta = pd.Series(beta)
    t_lvl, c_lvl = _window(tw, origin, level_k).mean(axis=1), _window(cw, origin, level_k).mean(axis=1)
    tp = temp.pivot(index="city_id", columns="year", values="predicted_value")
    cdd = tp.sub(t_lvl, axis=0).mul(beta, axis=0).add(c_lvl, axis=0).clip(lower=0)
    return _long({"cdd": cdd})


_CURVES = {}


def load_cdd_curves(data_dir: Path = DATA) -> pd.DataFrame:
    """(city_id, year) x shift table from prepare_data.py: CDD of that year with every day s °F warmer."""
    if data_dir not in _CURVES:
        c = pd.read_parquet(data_dir / "cdd_shift_curves.parquet")
        if c["Year"].max() > CUTOFF:
            raise RuntimeError(f"Data rule: found weather dated after {CUTOFF}; refusing to continue.")
        _CURVES[data_dir] = c.rename(columns={"ID": "city_id", "Year": "year"}).set_index(["city_id", "year"])
    return _CURVES[data_dir]


def f_cdd_delta(panel, meta, origin, years, *, temp_params=None, k=15):
    """Delta-method CDD: replay the daily weather of the city's last `k` valid CDD years,
    with every day warmed by (temperature forecast - mean temperature of those years), and
    average the resulting CDD. Captures that a city with many days just under 65 °F gains
    more CDD per degree than one that is rarely warm."""
    curves = load_cdd_curves()
    shifts = np.array([float(c[1:]) for c in curves.columns])
    temp = f_level_trend(panel, meta, origin, years, metrics=["annual_avg_temp"],
                         params={"annual_avg_temp": temp_params} if temp_params else None)
    tp = temp.pivot(index="city_id", columns="year", values="predicted_value")
    cw, tw = wide(panel, "cdd", origin), wide(panel, "annual_avg_temp", origin)
    out = {}
    for city in tp.index:
        yrs = [y for y in cw.columns[::-1] if pd.notna(cw.loc[city, y]) and pd.notna(tw.loc[city, y])
               and (city, y) in curves.index][:k]
        if not yrs:
            out[city] = pd.Series(np.nan, index=tp.columns)
            continue
        C = curves.loc[[(city, y) for y in yrs]].to_numpy()
        base_t = tw.loc[city, yrs].mean()
        out[city] = pd.Series({y: np.mean([np.interp(tp.loc[city, y] - base_t, shifts, row) for row in C])
                               for y in tp.columns})
    return _long({"cdd": pd.DataFrame(out).T})


def blend(*preds: pd.DataFrame) -> pd.DataFrame:
    """Equal-weight average of forecasts (rows matched on city_id, year, metric)."""
    return (pd.concat(preds).groupby(["city_id", "year", "metric"], as_index=False)["predicted_value"].mean())


# --------------------------------------------------------------------------- scoring

def score(pred: pd.DataFrame, panel: pd.DataFrame, baseline: pd.DataFrame, origin: int) -> dict:
    """Challenge score: S_m = 1 - MAE_model / MAE_baseline per metric, scored only on valid
    truth rows of city-metrics with >= 10 valid years in the 30 years before origin."""
    truth = panel[METRICS].stack().rename("truth").reset_index().rename(columns={"level_2": "metric"})
    hist = panel[(panel.index.get_level_values("year") > origin - 30) & (panel.index.get_level_values("year") <= origin)]
    counts = hist[METRICS].notna().groupby("city_id").sum().stack().rename("n").reset_index().rename(columns={"level_1": "metric"})
    df = (pred.merge(truth, on=["city_id", "year", "metric"])
              .merge(baseline.rename(columns={"predicted_value": "base"}), on=["city_id", "year", "metric"])
              .merge(counts, on=["city_id", "metric"]))
    df = df[df["n"] >= 10].dropna(subset=["truth", "predicted_value", "base"])
    res = {}
    for m in METRICS:
        d = df[df.metric == m]
        mae, mae_b = (d.predicted_value - d.truth).abs().mean(), (d.base - d.truth).abs().mean()
        res[m] = 1 - mae / mae_b
        res[m + "_mae"] = mae
    res["score"] = 100 * max(0.0, np.mean([res[m] for m in METRICS]))
    return res


def backtest(forecaster, panel, meta, origins, horizon=7, **kw) -> pd.DataFrame:
    rows = []
    for o in origins:
        years = list(range(o + 1, o + 1 + horizon))
        base = f_mean(panel, meta, o, years, k=30)
        pred = forecaster(panel, meta, o, years, **kw)
        rows.append({"origin": o, **score(pred, panel, base, o)})
    return pd.DataFrame(rows)