File size: 7,389 Bytes
05707f0 | 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 | """Table of the two exploratory extensions: real-time forecast shrinkage and U.S. state variables.
Panel A compares each learned model with its "_rs" version (forecast deviation scaled by a
slope estimated inside each training window) on capped value-weighted and equal-weighted
factors. Panel B compares models with and without the U.S. state variables on the sample
where all of them exist (feature months from 1990-01, so out of sample 2001-2025).
python scripts/extensions.py
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import numpy as np
import pandas as pd
MODELS = [("ols", "OLS"), ("enet", "Elastic net"), ("gbrt", "GBRT")]
SCOPES = ("own", "dm_only")
def _load(run: Path) -> tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]:
if json.loads((run / "run_manifest.json").read_text()).get("provenance") != "JKP":
raise SystemExit(f"{run} is not a real-data run")
r2 = pd.read_csv(run / "r2_oos.csv").set_index(["model", "scope"])
port = pd.read_csv(run / "portfolios.csv").set_index("strategy")
preds = pd.read_parquet(run / "predictions.parquet", columns=["model", "scope", "target_month", "scale"])
return r2, port, preds
def _cells(r2: pd.DataFrame, port: pd.DataFrame, model: str) -> dict:
row = {}
for s in SCOPES:
row[f"r2_{s}"] = r2.loc[(model, s), "r2_vs_shrink_pct"]
row[f"t_{s}"] = r2.loc[(model, s), "cw_t_shrink"]
row[f"sr_{s}"] = port.loc[f"{model}:{s}", "sharpe_net"]
return row
def _mean_scale(preds: pd.DataFrame, model: str, scope: str) -> float:
g = preds[(preds["model"] == model) & (preds["scope"] == scope)]
return float(g.groupby(g["target_month"].dt.year)["scale"].first().mean())
def _f(x: float, d: int = 2) -> str:
return "--" if pd.isna(x) else f"{x:.{d}f}".replace("-", "$-$")
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--rs", default="reports/jkp_em_rs")
ap.add_argument("--rs-ew", default="reports/jkp_em_ew_rs")
ap.add_argument("--no-macro", default="reports/jkp_em_from1990")
ap.add_argument("--macro", default="reports/jkp_em_macro")
ap.add_argument("--out", default="reports/robustness/extensions.csv")
ap.add_argument("--tex", default="paper/tables/extensions.tex")
a = ap.parse_args()
rows = []
for panel, label, run in (("A", "Capped value-weighted factors", a.rs), ("A", "Equal-weighted factors", a.rs_ew)):
r2, port, preds = _load(Path(run))
for m, name in MODELS:
rows.append({"panel": panel, "block": label, "model": name, "variant": "As estimated", **_cells(r2, port, m)})
rows.append({"panel": panel, "block": label, "model": name, "variant": "Real-time shrinkage",
**_cells(r2, port, f"{m}_rs"), "scale_own": _mean_scale(preds, f"{m}_rs", "own"),
"scale_dm_only": _mean_scale(preds, f"{m}_rs", "dm_only")})
rows.append({"panel": panel, "block": label, "model": "Static portfolio", "variant": "",
"sr_own": port.loc["static_ew", "sharpe_net"]})
rows.append({"panel": panel, "block": label, "model": "Shrinkage-mean portfolio", "variant": "",
"sr_own": port.loc["shrink_mean:rule", "sharpe_net"]})
base, with_macro = _load(Path(a.no_macro)), _load(Path(a.macro))
for m, name in MODELS:
rows.append({"panel": "B", "block": "From 1990", "model": name, "variant": "Without state variables",
**_cells(base[0], base[1], m)})
rows.append({"panel": "B", "block": "From 1990", "model": name, "variant": "With state variables",
**_cells(with_macro[0], with_macro[1], m)})
rows.append({"panel": "B", "block": "From 1990", "model": "Static portfolio", "variant": "",
"sr_own": base[1].loc["static_ew", "sharpe_net"]})
rows.append({"panel": "B", "block": "From 1990", "model": "Shrinkage-mean portfolio", "variant": "",
"sr_own": base[1].loc["shrink_mean:rule", "sharpe_net"]})
df = pd.DataFrame(rows)
Path(a.out).parent.mkdir(parents=True, exist_ok=True)
df.to_csv(a.out, index=False)
lines = [r"\begin{table}[!ht]", r"\centering\footnotesize\setlength{\tabcolsep}{4pt}",
r"\caption{Exploratory extensions}", r"\label{tab:extensions}",
r"\begin{tabular}{llrrrrrr}", r"\toprule",
r" & & \multicolumn{3}{c}{EM-trained} & \multicolumn{3}{c}{DM-trained} \\",
r"\cmidrule(lr){3-5}\cmidrule(lr){6-8}",
r"Model & Forecast & $R^2_{OS}$ (\%) & CW $t$ & SR net & $R^2_{OS}$ (\%) & CW $t$ & SR net \\"]
titles = {"A": "Panel A: real-time forecast shrinkage, out of sample 2000--2025",
"B": "Panel B: U.S. state variables, feature months from 1990, out of sample 2001--2025"}
for panel in ("A", "B"):
lines += [r"\midrule", rf"\multicolumn{{8}}{{l}}{{\textit{{{titles[panel]}}}}} \\"]
for block, g in df[df["panel"] == panel].groupby("block", sort=False):
if panel == "A":
lines.append(rf"\multicolumn{{8}}{{l}}{{{block}}} \\")
prev = None
for _, r in g.iterrows():
if r["variant"] == "":
lines.append(f"{r['model']} & & & & {_f(r['sr_own'])} & & & \\\\")
continue
name = "" if r["model"] == prev else r["model"]
prev = r["model"]
lines.append(f"{name} & {r['variant']} & " + " & ".join(
[_f(r[f'r2_{s}'], 3) + " & " + _f(r[f't_{s}']) + " & " + _f(r[f'sr_{s}']) for s in SCOPES]) + r" \\")
lines.append(r"\addlinespace")
sc = df[df["variant"] == "Real-time shrinkage"][["scale_own", "scale_dm_only"]].to_numpy()
lines += [r"\bottomrule", r"\end{tabular}", "", r"\vspace{4pt}",
r"\begin{minipage}{0.97\linewidth}\footnotesize",
r"\textit{Notes:} $R^2_{OS}$ and Clark--West $t$ against the shrinkage mean; SR net is the net Sharpe "
r"ratio of the timing portfolio at 20 basis points per unit of turnover (for the static and "
r"shrinkage-mean rows, of those portfolios). Real-time shrinkage multiplies each forecast's "
r"deviation from the shrinkage mean by a slope in $[0,1]$ estimated inside the training window: "
r"the model is fitted on labels before the window's last 36 months and the slope of realized on "
r"forecast deviations over those months is the scale; the model is then refitted on the whole "
rf"window. Average yearly scales range from {np.nanmin(sc):.2f} to {np.nanmax(sc):.2f}. "
r"State variables: log VIX, the U.S. 10-year minus 3-month Treasury spread and the 12-month log "
r"change of the broad trade-weighted dollar (FRED), dated at the end of the feature month; both "
r"rows of Panel B use the same sample. Both extensions were chosen after seeing the baseline "
r"results and are exploratory.",
r"\end{minipage}", r"\end{table}"]
Path(a.tex).write_text("\n".join(lines) + "\n")
with pd.option_context("display.width", 250, "display.max_columns", 20):
print(df.drop(columns=["panel"]).round(3).to_string(index=False))
if __name__ == "__main__":
main()
|