SPB Git
spboucher.ai

Repository Public Research preview

spbwp7_uqo

Commit

Stage E exhibits (Agent 9): nine manuscript figures from the v2 results with float fragments — fig_coverage, fig_rq2_loss_ratios, fig_rq2_subperiod, fig_rq3_headline, fig_rq3_series, fig_irf_bands (core); fig_spec_curve, fig_magnet, fig_rq5_clipping (appendix); scripts/33 owns them, scripts/12 and 17 no longer overwrite (WP07-ECON-18, ECON-13, THEO-10, EDIT-15, ADVE-17, cluster H 'no figures')

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed · Nov 4, 2026 parent ed79566
30 changed files +513 −1 Browse files
30 changed files
Unified Split
added figures/fig_coverage.pdf +0 −0

Binary file not shown.

added figures/fig_coverage.png +0 −0

Binary file not shown.

modified figures/fig_irf_bands.pdf +0 −0

Binary file not shown.

modified figures/fig_irf_bands.png +0 −0

Binary file not shown.

added figures/fig_magnet.pdf +0 −0

Binary file not shown.

added figures/fig_magnet.png +0 −0

Binary file not shown.

added figures/fig_rq2_loss_ratios.pdf +0 −0

Binary file not shown.

added figures/fig_rq2_loss_ratios.png +0 −0

Binary file not shown.

added figures/fig_rq2_subperiod.pdf +0 −0

Binary file not shown.

added figures/fig_rq2_subperiod.png +0 −0

Binary file not shown.

added figures/fig_rq3_headline.pdf +0 −0

Binary file not shown.

added figures/fig_rq3_headline.png +0 −0

Binary file not shown.

added figures/fig_rq3_series.pdf +0 −0

Binary file not shown.

added figures/fig_rq3_series.png +0 −0

Binary file not shown.

added figures/fig_rq5_clipping.pdf +0 −0

Binary file not shown.

added figures/fig_rq5_clipping.png +0 −0

Binary file not shown.

modified figures/fig_spec_curve.pdf +0 −0

Binary file not shown.

added figures/fig_spec_curve.png +0 −0

Binary file not shown.

added paper/figures/fig_coverage.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_coverage from results/descriptive_coverage_complete.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_coverage.pdf}
5 +\caption{Coverage of complete-predictor ticker-days by year and asset group, 2010--2025. A ticker-day is complete when the ten option-implied predictors, $RV_t$, $RV^{(w)}_t$ and the next-day return are all observed; the 90-day tenor (needed for the term slope) is the binding constraint. Lines: share of complete ticker-days within each asset group (markers and line styles distinguish the groups in grayscale). Over the full panel 119{,}081 of 264{,}383 ticker-days (45.0\%) are complete; the share rises from 15\% in 2010 to 63\% in 2025. Source: \texttt{results/descriptive\_coverage\_complete.csv} (package L1-E3).}
6 +\label{fig:coverage}
7 +\end{figure}
added paper/figures/fig_irf_bands.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_irf_bands from results/irf_bands.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_irf_bands.pdf}
5 +\caption{Orthogonalized impulse responses of the bivariate VAR(5) in ATM implied volatility and realized variance (standardized levels), averaged across the 68 tickers with at least 1{,}000 complete observations. Lines: cross-ticker mean response to a one-standard-deviation shock under the two recursive orderings (solid, circles: RV ordered first, so that RV shocks move IV contemporaneously but not the reverse; dashed, squares: IV first); shaded bands (plain / hatched): cross-ticker average of the per-ticker residual-bootstrap 5th and 95th percentiles (500 draws, seed 42). Under the RV-first ordering the RV response to an IV shock peaks at 0.150 SD at horizon 1 (band [0.126, 0.173]). Variance decompositions and stationarity diagnostics are in Table~\ref{tab:fevd_v2}. Source: \texttt{results/irf\_bands.csv} (package L1-E8).}
6 +\label{fig:irf_bands}
7 +\end{figure}
added paper/figures/fig_magnet.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_magnet from results/rq4_price_magnet_v2.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_magnet.pdf}
5 +\caption{Price-magnet test on the shipped panel (10{,}480 ticker-days of eight stocks, 2010--2025, open within 0.1\%--10\% of the maximum-open-interest strike among options expiring within 30 days). Circles: observed share of days on which the close is relatively closer to that strike than the open, with $\pm1.96$ date-clustered standard errors; squares: the permutation null that re-pairs each (day, distance, strike side) with the open-to-close move of a random other day of the same ticker (1000 draws, seed 42), mean and 95\% interval; crosses: the mirror-strike placebo. The dotted line is the 50\% benchmark of version~1.2, which is not the correct null because a driftless move reaches a strike at distance $d$ only if it has the right sign and does not overshoot. Over all ticker-days the observed share is 47.01\% against a null of 46.45\% [45.6, 47.3] (permutation $p$ = 0.20; minimum detectable difference 2.0 pp at 80\% power). Source: \texttt{results/rq4\_price\_magnet\_v2.csv} (Table~\ref{tab:magnet_v2}).}
6 +\label{fig:magnet}
7 +\end{figure}
added paper/figures/fig_rq2_loss_ratios.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_rq2_loss_ratios from results/rq2_loss_ratio_ci.csv, results/rq2_diebold_mariano_v2.csv, results/rq2_oos_summary_v2.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_rq2_loss_ratios.pdf}
5 +\caption{Out-of-sample loss of HAR-RV $+$ IV surface relative to HAR-RV, ticker by ticker (67 tickers, rolling 500/250-day windows, training-window winsorization, 2010--2025). Each point is the ratio of the summed out-of-sample loss of the augmented model to that of HAR-RV (below one: the IV surface lowers the loss); vertical segments are 95\% stationary-bootstrap intervals of the ratio (1{,}000 draws, mean block 10 days); marker shape and colour give the sign of the per-ticker Diebold--Mariano statistic (Newey--West variance, $h$ lags). For $RV_{t+1}$ the median MSE ratio is 1.001 (6 tickers significantly better, 19 worse) and the median QLIKE ratio is 1.284 (0 better, 38 worse); the panel DM statistics (pooled loss differential, date-aggregated HAC variance) are -0.94 and -6.49. The Diebold--Mariano statistic is biased toward ``worse'' for the nested augmented model (Table~\ref{tab:rq2_null_mc}); Clark--West statistics are in Table~\ref{tab:rq2_cw}. Sources: \texttt{results/rq2\_loss\_ratio\_ci.csv}, \texttt{rq2\_diebold\_mariano\_v2.csv}, \texttt{rq2\_oos\_summary\_v2.csv}.}
6 +\label{fig:rq2_loss_ratios}
7 +\end{figure}
added paper/figures/fig_rq2_subperiod.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_rq2_subperiod from results/subperiod_results_v2.csv, results/rq2_subperiod_oos.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_rq2_subperiod.pdf}
5 +\caption{Stability of the realized-variance results across the nine subperiods of the sample. Panel (a): adjusted in-sample $R^2$ of HAR-RV and of HAR-RV $+$ IV surface for $RV_{t+1}$, re-estimated within each subperiod on the common sample (per-ticker winsorization); the relative gain ranges from 4.3\% to 114.1\% and, because the models are nested, cannot be negative in sample. Panel (b): median across tickers (at least 30 forecasts in the subperiod, at least five tickers) of the ratio of the out-of-sample loss of the augmented model to that of HAR-RV, from the rolling 500/250 forecasts with training-window winsorization; the 2010--12 subperiod has forecasts for one ticker only and is not shown. For the IV-surface model the median MSE ratio ranges from 0.958 to 1.116 and the median QLIKE ratio from 0.956 to 1.638. Sources: \texttt{results/subperiod\_results\_v2.csv}, \texttt{rq2\_subperiod\_oos.csv} (package L1-B3).}
6 +\label{fig:rq2_subperiod}
7 +\end{figure}
added paper/figures/fig_rq3_headline.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_rq3_headline from results/rq3_headline.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_rq3_headline.pdf}
5 +\caption{Implied--realized correlation ratio as a predictor of market stress: the nine headline coefficients with episode-bootstrap intervals. Linear probability models of the indicator that a post-origin stress event (definition S2: VIX$_{t+j}>25$ or VIX$_{t+j}/$VIX$_t-1>20\%$ for some $j\le h$) occurs within $h$ trading days, on the standardized ratio and the extended controls, rebuilt series 2010--2025 (3{,}668 days at $h=5$). Markers: coefficient in percentage points of probability per standard deviation of the transformed ratio; bars: 95\% interval from resampling the maximal runs of equal target value (stress episodes and calm spells) with replacement, 1{,}000 draws; labels: Holm step-down $p$-values over the nine tests. For the log ratio the coefficients are 1.77, 4.20 and 3.78 pp (Newey--West($h$) $t$ of 1.55, 2.50, 1.73); no interval excludes zero and the smallest Holm $p$-value in the figure is 0.005 (raw ratio, $h=10$; Romano--Wolf $p$ = 0.12). Source: \texttt{results/rq3\_headline.csv} (Table~\ref{tab:rq3_headline}).}
6 +\label{fig:rq3_headline}
7 +\end{figure}
added paper/figures/fig_rq3_series.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_rq3_series from results/correlation_divergence_v2.parquet — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_rq3_series.pdf}
5 +\caption{Implied and realized correlation among S\&P~500 constituents, 2010--2025 (3{,}678 trading days, series rebuilt from the shipped parquets by stage~1b of \texttt{scripts/04}, 29 constituents, days with fewer than 20 constituents dropped). Panel (a): the CBOE-style implied correlation from the ATM implied variances of SPX and of the constituents (eq.~(\ref{eq:implcorr}), clipped to $[0,1]$) and the 22-day pairwise realized correlation; panel (b): the log ratio that enters the stress regressions (mean $-$0.252, standard deviation 0.444); panel (c): the VIX close from the complete daily series of the FinData OS lake. Shaded: days on which the VIX closes above 25 (13.3\% of days), the level component of the post-origin stress definition. Source: \texttt{data/processed/correlation\_divergence\_v2.parquet} (regenerable; diagnostics in Table~\ref{tab:rq3_diag}).}
6 +\label{fig:rq3_series}
7 +\end{figure}
added paper/figures/fig_rq5_clipping.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_rq5_clipping from results/rq5_clipping_variant.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_rq5_clipping.pdf}
5 +\caption{Dependence of the RQ5 benchmark comparison on the treatment of the test-year regressors (SPX, next-day realized variance, rolling annual re-estimation 2016--2025, FinData OS lake data, \texttt{scripts/32}). Lines: out-of-sample $R^2$ of the HAR-RV model by test year (vertical axis truncated below the lowest HAR-RV value; off-scale values of the VIX benchmark are marked with their value) when the test-year regressors are clipped at the training window's 1\%/99\% quantiles (the scheme of the RQ2 design), left unclipped, or when the model is estimated in logs; crosses: the VIX benchmark. Clipping caps $RV_t$ in 2020 far below its realized level: the HAR-RV $R^2_{OOS}$ for 2020 is 0.058 clipped against 0.449 unclipped, and over 2016--2025 it is 0.158 clipped against 0.465 unclipped, versus 0.347 for the VIX benchmark (Diebold--Mariano $t$ of the unclipped HAR-RV against the VIX: 1.28). Source: \texttt{results/rq5\_clipping\_variant.csv} (Table~\ref{tab:rq5_v2}, Panel C).}
6 +\label{fig:rq5_clipping}
7 +\end{figure}
added paper/figures/fig_spec_curve.tex +7 −0
@@ -0,0 +1,7 @@
1 +% Generated by scripts/33_make_exhibit_figures.py::fig_spec_curve from results/rq1_spec_curve.csv — do not edit; regenerate with `make tables` (figure).
2 +\begin{figure}[tbp]
3 +\centering
4 +\includegraphics[width=\textwidth]{fig_spec_curve.pdf}
5 +\caption{Specification curve of the 25-delta skew coefficient in the return regressions (400 specifications: winsorization cutoff $\times$ standard-error estimator $\times$ sample $\times$ horizon $\times$ controls $\times$ return basis; Table~\ref{tab:spec_curve}). Panel (a): $t$-statistics sorted in increasing order, marker and colour by standard-error estimator; dashed lines at $\pm1.96$, dotted at $\pm3.02$. Panel (b): the specification behind each point. The median $t$ is 1.93; $|t|>1.96$ in 50.0\% and $|t|>3.02$ in 18.5\% of specifications; among the Driscoll--Kraay specifications the share with $|t|>3.02$ is 5.0\%. Source: \texttt{results/rq1\_spec\_curve.csv} (package L1-C5).}
6 +\label{fig:spec_curve}
7 +\end{figure}
modified scripts/12_make_figures.py +2 −0
@@ -118,6 +118,8 @@ def fig_irf_bands() -> None:
118 118 to an RV shock under both recursive orderings (levels specification), from
119 119 results/irf_bands.csv written by scripts/11. New output fig_irf_bands; the
120 120 legacy fig_irf (reduced-form IRFs, 20 tickers) is unchanged."""
121 + print(" [fig_irf_bands: the manuscript version is drawn by scripts/33_make_exhibit_figures.py (Stage E design system); skipped here]")
122 + return
121 123 path = config.RESULTS_DIR / "irf_bands.csv"
122 124 if not path.exists():
123 125 print(" [fig_irf_bands skipped: results/irf_bands.csv missing — run scripts/11]")
modified scripts/17_specification_curve.py +6 −1
@@ -131,7 +131,12 @@ def main():
131 131 share_196=('abs_t_gt_196', 'mean'), share_302=('abs_t_gt_302', 'mean'))
132 132 print(f"\n by {key}:\n" + g.round(3).to_string())
133 133
134 − # Figure
134 + # Figure: the manuscript version (figures/fig_spec_curve.pdf + paper/figures/fig_spec_curve.tex) is drawn by
135 + # scripts/33_make_exhibit_figures.py from results/rq1_spec_curve.csv (Stage E design system); the legacy drawing is kept below but not written.
136 + if True:
137 + print("\n [fig_spec_curve: drawn by scripts/33_make_exhibit_figures.py]")
138 + print("\nSPECIFICATION CURVE COMPLETE.")
139 + return
135 140 import matplotlib
136 141 matplotlib.use('Agg')
137 142 import matplotlib.pyplot as plt
added scripts/33_make_exhibit_figures.py +442 −0
@@ -0,0 +1,442 @@
1 +#!/usr/bin/env python3
2 +# =============================================================================
3 +# Author: Simon-Pierre Boucher
4 +# Contact: contact@spboucher.ai
5 +# =============================================================================
6 +"""Step 33 — Manuscript figures (Stage E, 2026-11-04; Part X.2 standard).
7 +
8 +Every figure is drawn from a validated result file of the Stage D revision
9 +(results/*.csv, or the regenerable processed parquets for the RQ3 series),
10 +in one design system (src/wp7/figstyle.py: Okabe-Ito palette, markers and
11 +line styles for grayscale, vector PDF at \\textwidth, embedded serif face),
12 +and comes with a float fragment paper/figures/<name>.tex whose caption
13 +quotes only numbers read from the same files.
14 +
15 +Core figures
16 + fig_coverage share of complete ticker-days by year and asset group
17 + fig_rq2_loss_ratios per-ticker OOS loss ratios with bootstrap intervals
18 + fig_rq2_subperiod in-sample adj. R^2 and OOS loss ratios by subperiod
19 + fig_rq3_headline RQ3 coefficients with episode-bootstrap intervals
20 + fig_rq3_series implied vs realized correlation, log ratio, VIX
21 + fig_irf_bands orthogonalized IRFs with residual-bootstrap bands
22 +Appendix figures
23 + fig_spec_curve specification curve of the 25-delta skew
24 + fig_magnet price-magnet shares against the permutation null
25 + fig_rq5_clipping RQ5 R^2_OOS by year under the three clipping rules
26 +
27 +Inputs : results/descriptive_coverage_complete.csv, rq2_loss_ratio_ci.csv,
28 + rq2_diebold_mariano_v2.csv, rq2_oos_summary_v2.csv, subperiod_results_v2.csv,
29 + rq2_subperiod_oos.csv, rq3_headline.csv, irf_bands.csv, rq1_spec_curve.csv,
30 + rq4_price_magnet_v2.csv, rq5_clipping_variant.csv,
31 + data/processed/correlation_divergence_v2.parquet (regenerable, scripts/04)
32 +Outputs: figures/<name>.pdf (+ .png), paper/figures/<name>.tex
33 +"""
34 +
35 +import warnings
36 +
37 +import numpy as np
38 +import pandas as pd
39 +from matplotlib.lines import Line2D
40 +from matplotlib.patches import Patch
41 +
42 +import _bootstrap # noqa: F401
43 +from wp7 import config
44 +from wp7 import figstyle as fs
45 +from wp7.tables import SUBPERIOD_LABEL, f, figure, num
46 +
47 +warnings.filterwarnings('ignore')
48 +R = config.RESULTS_DIR
49 +P = config.DATA_PROCESSED
50 +
51 +
52 +def _txt(x, d=2):
53 + """Plain-text number for captions (LaTeX minus)."""
54 + return f(x, d)
55 +
56 +
57 +# ----------------------------------------------------------------------------
58 +# core: data coverage
59 +# ----------------------------------------------------------------------------
60 +def fig_coverage():
61 + c = pd.read_csv(R / "descriptive_coverage_complete.csv")
62 + g = c.groupby(['group', 'year']).agg(n=('n_rows', 'sum'), k1=('n_complete_rq1', 'sum'), k2=('n_complete_rq2', 'sum')).reset_index()
63 + a = c.groupby('year').agg(n=('n_rows', 'sum'), k1=('n_complete_rq1', 'sum'), k2=('n_complete_rq2', 'sum')).reset_index()
64 + fig, ax = fs.new(width=6.5, height=2.9)
65 + for i, grp in enumerate(['Stocks', 'ETFs', 'Indices']):
66 + s = g[g['group'] == grp]
67 + ax.plot(s['year'], s['k1'] / s['n'] * 100, color=fs.SERIES[i], marker=fs.MARKERS[i], ls=fs.LINESTYLES[i], label=f"{grp} (n = {c[c['group']==grp]['ticker'].nunique()})")
68 + ax.plot(a['year'], a['k1'] / a['n'] * 100, color=fs.INK, marker='x', ls='-', lw=1.0, label='All 69 underlyings')
69 + ax.set_ylim(0, 100)
70 + ax.set_xlabel('Year')
71 + ax.set_ylabel('Complete ticker-days (% of ticker-days)')
72 + ax.set_xticks(sorted(a['year'].unique()))
73 + ax.tick_params(axis='x', labelrotation=45)
74 + ax.legend(loc='upper left', ncol=2)
75 + fs.save(fig, 'fig_coverage')
76 + tot_n, tot_k = int(a['n'].sum()), int(a['k1'].sum())
77 + first, last = a.iloc[0], a.iloc[-1]
78 + cap = (rf"Coverage of complete-predictor ticker-days by year and asset group, 2010--2025. A ticker-day is complete when the ten option-implied "
79 + rf"predictors, $RV_t$, $RV^{{(w)}}_t$ and the next-day return are all observed; the 90-day tenor (needed for the term slope) is the binding "
80 + rf"constraint. Lines: share of complete ticker-days within each asset group (markers and line styles distinguish the groups in grayscale). "
81 + rf"Over the full panel {num(tot_k)} of {num(tot_n)} ticker-days ({_txt(tot_k/tot_n*100,1)}\%) are complete; the share rises from "
82 + rf"{_txt(first['k1']/first['n']*100,0)}\% in {int(first['year'])} to {_txt(last['k1']/last['n']*100,0)}\% in {int(last['year'])}. "
83 + rf"Source: \texttt{{results/descriptive\_coverage\_complete.csv}} (package L1-E3).")
84 + figure('fig_coverage', cap, 'fig:coverage')
85 +
86 +
87 +# ----------------------------------------------------------------------------
88 +# core: RQ2 per-ticker loss ratios
89 +# ----------------------------------------------------------------------------
90 +def fig_rq2_loss_ratios():
91 + lr = pd.read_csv(R / "rq2_loss_ratio_ci.csv")
92 + dm = pd.read_csv(R / "rq2_diebold_mariano_v2.csv")
93 + summ = pd.read_csv(R / "rq2_oos_summary_v2.csv")
94 + summ = summ[(summ['scheme'] == 'training-window winsorization') & (summ['subset'] == 'all')]
95 + fig, axes = fs.new(width=6.5, height=4.6, nrows=2, ncols=2, sharex=False)
96 + panels = [('1-Day RV', 'mse', '(a) MSE, target $RV_{t+1}$'), ('1-Day RV', 'qlike', '(b) QLIKE, target $RV_{t+1}$'),
97 + ('5-Day RV', 'mse', r'(c) MSE, target $\sum_{j=1}^{5} RV_{t+j}$'), ('5-Day RV', 'qlike', r'(d) QLIKE, target $\sum_{j=1}^{5} RV_{t+j}$')]
98 + col_better, col_worse, col_ns = fs.OI['blue'], fs.OI['vermillion'], fs.OI['grey']
99 + for ax, (tgt, loss, title) in zip(axes.ravel(), panels):
100 + s = lr[lr['target'] == tgt].merge(dm[dm['target'] == tgt][['ticker', 'dm_mse', 'dm_qlike']], on='ticker')
101 + s = s.sort_values(f'{loss}_ratio').reset_index(drop=True)
102 + x = np.arange(len(s))
103 + d = s[f'dm_{loss}']
104 + colors = np.where(d > 1.96, col_better, np.where(d < -1.96, col_worse, col_ns))
105 + markers = np.where(d > 1.96, 'o', np.where(d < -1.96, 's', 'o'))
106 + ax.vlines(x, s[f'{loss}_ratio_p025'], s[f'{loss}_ratio_p975'], color=colors, lw=0.8, alpha=0.8)
107 + for mk in ('o', 's'):
108 + sel = markers == mk
109 + ax.scatter(x[sel], s.loc[sel, f'{loss}_ratio'], s=12, marker=mk, c=colors[sel], edgecolor=fs.INK, linewidth=0.3, zorder=3)
110 + ax.axhline(1.0, color=fs.INK, lw=0.8, ls='--')
111 + fs.panel_label(ax, title)
112 + ax.set_xlabel('Tickers ordered by loss ratio')
113 + ax.set_ylabel('Loss ratio, augmented / HAR-RV')
114 + ax.set_xticks([])
115 + q = summ[(summ['target'] == tgt) & (summ['loss'] == loss)].iloc[0]
116 + ax.text(0.02, 0.97, f"median {q['median_ratio']:.3f}; DM $>1.96$: {int(q['n_dm_gt_1.96'])}, $<-1.96$: {int(q['n_dm_lt_-1.96'])} of {int(q['n_tickers'])}\n"
117 + f"Holm 5%: {int(q['holm_pos'])} / {int(q['holm_neg'])}; panel DM $t$ = {q['panel_dm_t']:.2f}",
118 + transform=ax.transAxes, va='top', ha='left', fontsize=7.5)
119 + handles = [Line2D([], [], marker='o', color=col_better, ls='', markeredgecolor=fs.INK, markeredgewidth=0.3, label='DM $> 1.96$ (augmented model better)'),
120 + Line2D([], [], marker='s', color=col_worse, ls='', markeredgecolor=fs.INK, markeredgewidth=0.3, label='DM $< -1.96$ (augmented model worse)'),
121 + Line2D([], [], marker='o', color=col_ns, ls='', markeredgecolor=fs.INK, markeredgewidth=0.3, label='$|$DM$| \\leq 1.96$'),
122 + Line2D([], [], color=fs.INK2, lw=0.8, label='95% stationary-bootstrap interval')]
123 + fig.legend(handles=handles, loc='outside lower center', ncol=2)
124 + fs.save(fig, 'fig_rq2_loss_ratios')
125 + q1m = summ[(summ['target'] == '1-Day RV') & (summ['loss'] == 'mse')].iloc[0]
126 + q1q = summ[(summ['target'] == '1-Day RV') & (summ['loss'] == 'qlike')].iloc[0]
127 + n_t = int(q1m['n_tickers'])
128 + cap = (rf"Out-of-sample loss of HAR-RV $+$ IV surface relative to HAR-RV, ticker by ticker ({n_t} tickers, rolling 500/250-day windows, "
129 + rf"training-window winsorization, 2010--2025). Each point is the ratio of the summed out-of-sample loss of the augmented model to that of HAR-RV "
130 + rf"(below one: the IV surface lowers the loss); vertical segments are 95\% stationary-bootstrap intervals of the ratio (1{{,}}000 draws, mean block 10 days); "
131 + rf"marker shape and colour give the sign of the per-ticker Diebold--Mariano statistic (Newey--West variance, $h$ lags). "
132 + rf"For $RV_{{t+1}}$ the median MSE ratio is {q1m['median_ratio']:.3f} ({int(q1m['n_dm_gt_1.96'])} tickers significantly better, {int(q1m['n_dm_lt_-1.96'])} worse) "
133 + rf"and the median QLIKE ratio is {q1q['median_ratio']:.3f} ({int(q1q['n_dm_gt_1.96'])} better, {int(q1q['n_dm_lt_-1.96'])} worse); the panel DM statistics "
134 + rf"(pooled loss differential, date-aggregated HAC variance) are {q1m['panel_dm_t']:.2f} and {q1q['panel_dm_t']:.2f}. The Diebold--Mariano statistic "
135 + rf"is biased toward ``worse'' for the nested augmented model (Table~\ref{{tab:rq2_null_mc}}); Clark--West statistics are in Table~\ref{{tab:rq2_cw}}. "
136 + rf"Sources: \texttt{{results/rq2\_loss\_ratio\_ci.csv}}, \texttt{{rq2\_diebold\_mariano\_v2.csv}}, \texttt{{rq2\_oos\_summary\_v2.csv}}.")
137 + figure('fig_rq2_loss_ratios', cap, 'fig:rq2_loss_ratios')
138 +
139 +
140 +# ----------------------------------------------------------------------------
141 +# core: RQ2 by subperiod
142 +# ----------------------------------------------------------------------------
143 +def fig_rq2_subperiod():
144 + d = pd.read_csv(R / "subperiod_results_v2.csv")
145 + o = pd.read_csv(R / "rq2_subperiod_oos.csv")
146 + o = o[(o['target'] == '1-Day RV')]
147 + sps = list(config.SUBPERIODS)
148 + labels = [SUBPERIOD_LABEL[s] for s in sps]
149 + fig, axes = fs.new(width=6.5, height=3.1, nrows=1, ncols=2)
150 + ax = axes[0]
151 + har = [d[(d['subperiod'] == s) & (d['model'] == 'HAR-RV')]['adj_r2_model'].iloc[0] for s in sps]
152 + hiv = [d[(d['subperiod'] == s) & (d['model'] == 'HAR+IV')]['adj_r2_model'].iloc[0] for s in sps]
153 + x = np.arange(len(sps))
154 + ax.plot(x, har, color=fs.SERIES[0], marker=fs.MARKERS[0], ls=fs.LINESTYLES[0], label='HAR-RV')
155 + ax.plot(x, hiv, color=fs.SERIES[1], marker=fs.MARKERS[1], ls=fs.LINESTYLES[1], label='HAR-RV $+$ IV surface')
156 + ax.set_xticks(x); ax.set_xticklabels(labels, rotation=45, ha='right')
157 + ax.set_ylabel('Adjusted $R^2$, in sample ($RV_{t+1}$)')
158 + ax.set_ylim(0, max(hiv) * 1.15)
159 + ax.legend(loc='upper left')
160 + fs.panel_label(ax, '(a) In-sample fit by subperiod')
161 + ax = axes[1]
162 + for i, (comp, lab) in enumerate([('HAR-RV + IV surface / HAR-RV', 'HAR-RV $+$ IV surface'), ('HAR-RV + IV level / HAR-RV', r'HAR-RV $+$ $IV_{ATM,30d}$')]):
163 + for j, (loss, ll) in enumerate([('mse', 'MSE'), ('qlike', 'QLIKE')]):
164 + ys = []
165 + for s in sps:
166 + r = o[(o['subperiod'] == s) & (o['comparison'] == comp)]
167 + ys.append(r[f'{loss}_ratio_median'].iloc[0] if len(r) and r['n_tickers'].iloc[0] >= 5 else np.nan)
168 + k = i * 2 + j
169 + ax.plot(x, ys, color=fs.SERIES[k], marker=fs.MARKERS[k], ls=fs.LINESTYLES[k], label=f'{lab}, {ll}')
170 + ax.axhline(1.0, color=fs.INK, lw=0.8, ls='--')
171 + ax.set_xticks(x); ax.set_xticklabels(labels, rotation=45, ha='right')
172 + ax.set_ylabel('Median OOS loss ratio vs HAR-RV ($RV_{t+1}$)')
173 + ax.legend(loc='upper left', fontsize=7)
174 + fs.panel_label(ax, '(b) Out-of-sample loss ratios by subperiod')
175 + fs.save(fig, 'fig_rq2_subperiod')
176 + srf = o[o['comparison'] == 'HAR-RV + IV surface / HAR-RV']
177 + srf = srf[(srf['n_tickers'] >= 5) & (srf['subperiod'] != 'Full sample')]
178 + dr = [(h2 - h1) / h1 * 100 for h1, h2 in zip(har, hiv)]
179 + cap = (rf"Stability of the realized-variance results across the nine subperiods of the sample. Panel (a): adjusted in-sample $R^2$ of HAR-RV and of "
180 + rf"HAR-RV $+$ IV surface for $RV_{{t+1}}$, re-estimated within each subperiod on the common sample (per-ticker winsorization); the relative gain ranges from "
181 + rf"{_txt(min(dr),1)}\% to {_txt(max(dr),1)}\% and, because the models are nested, cannot be negative in sample. Panel (b): median across tickers "
182 + rf"(at least 30 forecasts in the subperiod, at least five tickers) of the ratio of the out-of-sample loss of the augmented model to that of HAR-RV, from the rolling "
183 + rf"500/250 forecasts with training-window winsorization; the 2010--12 subperiod has forecasts for one ticker only and is not shown. For the IV-surface model "
184 + rf"the median MSE ratio ranges from {srf['mse_ratio_median'].min():.3f} to {srf['mse_ratio_median'].max():.3f} and the median QLIKE ratio from "
185 + rf"{srf['qlike_ratio_median'].min():.3f} to {srf['qlike_ratio_median'].max():.3f}. "
186 + rf"Sources: \texttt{{results/subperiod\_results\_v2.csv}}, \texttt{{rq2\_subperiod\_oos.csv}} (package L1-B3).")
187 + figure('fig_rq2_subperiod', cap, 'fig:rq2_subperiod')
188 +
189 +
190 +# ----------------------------------------------------------------------------
191 +# core: RQ3 headline coefficients
192 +# ----------------------------------------------------------------------------
193 +def fig_rq3_headline():
194 + h = pd.read_csv(R / "rq3_headline.csv")
195 + trs = [('raw', 'Raw ratio'), ('log', 'Log ratio'), ('rank', 'Rank')]
196 + hz = ['5d', '10d', '20d']
197 + fig, ax = fs.new(width=6.5, height=3.0)
198 + for i, (tr, lab) in enumerate(trs):
199 + xs, ys, lo, hi, ph = [], [], [], [], []
200 + for j, z in enumerate(hz):
201 + r = h[(h['transform'] == tr) & (h['horizon'] == z)].iloc[0]
202 + xs.append(j + (i - 1) * 0.24); ys.append(r['coefficient_pp']); lo.append(r['boot_ci_lo_pp']); hi.append(r['boot_ci_hi_pp']); ph.append(r['p_holm'])
203 + xs, ys, lo, hi = map(np.array, (xs, ys, lo, hi))
204 + ax.errorbar(xs, ys, yerr=[ys - lo, hi - ys], fmt=fs.MARKERS[i], color=fs.SERIES[i], capsize=3, lw=1.1, label=lab, markeredgecolor=fs.INK, markeredgewidth=0.3)
205 + for xx, ll, p in zip(xs, lo, ph):
206 + ax.annotate(f"Holm\n$p$ = {p:.3f}", xy=(xx, ll), xytext=(0, -4), textcoords='offset points', ha='center', va='top', fontsize=6.8, color=fs.INK2)
207 + ax.axhline(0, color=fs.INK, lw=0.8)
208 + ax.set_xticks(range(3)); ax.set_xticklabels([f'$h = {z[:-1]}$ days' for z in hz])
209 + ax.set_ylabel('Stress probability, pp per SD of the ratio')
210 + ax.set_xlabel('Forecast horizon')
211 + fig.legend(loc='outside lower center', ncol=3, title='Transform of $\\hat\\rho_{impl}/\\hat\\rho_{real}$', title_fontsize=8)
212 + ax.margins(x=0.12, y=0.18)
213 + fs.save(fig, 'fig_rq3_headline')
214 + n = int(h['n_obs'].max()); lg = h[h['transform'] == 'log'].set_index('horizon')
215 + cap = (rf"Implied--realized correlation ratio as a predictor of market stress: the nine headline coefficients with episode-bootstrap intervals. "
216 + rf"Linear probability models of the indicator that a post-origin stress event (definition S2: VIX$_{{t+j}}>25$ or VIX$_{{t+j}}/$VIX$_t-1>20\%$ for some $j\le h$) "
217 + rf"occurs within $h$ trading days, on the standardized ratio and the extended controls, rebuilt series 2010--2025 ({num(n)} days at $h=5$). Markers: coefficient in "
218 + rf"percentage points of probability per standard deviation of the transformed ratio; bars: 95\% interval from resampling the maximal runs of equal target value "
219 + rf"(stress episodes and calm spells) with replacement, 1{{,}}000 draws; labels: Holm step-down $p$-values over the nine tests. For the log ratio the coefficients are "
220 + rf"{_txt(lg.loc['5d','coefficient_pp'])}, {_txt(lg.loc['10d','coefficient_pp'])} and {_txt(lg.loc['20d','coefficient_pp'])} pp "
221 + rf"(Newey--West($h$) $t$ of {_txt(lg.loc['5d','t_nw'])}, {_txt(lg.loc['10d','t_nw'])}, {_txt(lg.loc['20d','t_nw'])}); no interval excludes zero and the smallest "
222 + rf"Holm $p$-value in the figure is {h['p_holm'].min():.3f} (raw ratio, $h=10$; Romano--Wolf $p$ = {h.loc[h['p_holm'].idxmin(),'p_romano_wolf']:.2f}). "
223 + rf"Source: \texttt{{results/rq3\_headline.csv}} (Table~\ref{{tab:rq3_headline}}).")
224 + figure('fig_rq3_headline', cap, 'fig:rq3_headline')
225 +
226 +
227 +# ----------------------------------------------------------------------------
228 +# core: RQ3 series
229 +# ----------------------------------------------------------------------------
230 +def fig_rq3_series():
231 + path = P / "correlation_divergence_v2.parquet"
232 + if not path.exists():
233 + print(" [fig_rq3_series skipped: data/processed/correlation_divergence_v2.parquet missing — run scripts/04]")
234 + return
235 + d = pd.read_parquet(path).sort_values('trade_date')
236 + d['log_ratio'] = np.log(d['corr_ratio'].where(d['corr_ratio'] > 0))
237 + fig, axes = fs.new(width=6.5, height=5.6, nrows=3, ncols=1, sharex=True)
238 + hi_vix = d['vix_close'] > 25
239 + for ax in axes:
240 + ax.fill_between(d['trade_date'], 0, 1, where=hi_vix, transform=ax.get_xaxis_transform(), color=fs.OI['grey'], alpha=0.25, lw=0)
241 + ax = axes[0]
242 + ax.plot(d['trade_date'], d['implied_corr'], color=fs.SERIES[0], lw=0.9, ls='-', label='Implied correlation (29 constituents, CBOE-style)')
243 + ax.plot(d['trade_date'], d['realized_corr'], color=fs.SERIES[1], lw=0.9, ls='--', label='Realized correlation (22-day pairwise)')
244 + ax.set_ylabel('Correlation')
245 + ax.set_ylim(0, 1)
246 + ax.legend(loc='upper right', ncol=2)
247 + fs.panel_label(ax, '(a) Implied and realized correlation among S&P 500 constituents')
248 + ax = axes[1]
249 + ax.plot(d['trade_date'], d['log_ratio'], color=fs.SERIES[2], lw=0.8)
250 + ax.axhline(0, color=fs.INK, lw=0.7)
251 + ax.set_ylabel(r'$\ln(\hat\rho_{impl}/\hat\rho_{real})$')
252 + fs.panel_label(ax, '(b) Log ratio of implied to realized correlation')
253 + ax = axes[2]
254 + ax.plot(d['trade_date'], d['vix_close'], color=fs.INK, lw=0.8)
255 + ax.axhline(25, color=fs.SERIES[1], lw=0.8, ls='--')
256 + ax.text(d['trade_date'].iloc[len(d) // 2], 27, 'VIX = 25 (stress threshold)', fontsize=7.5, color=fs.SERIES[1], va='bottom', ha='center',
257 + bbox=dict(facecolor='white', edgecolor='none', pad=0.6))
258 + ax.set_ylabel('VIX close')
259 + ax.set_xlabel('Trading day')
260 + fs.panel_label(ax, '(c) VIX')
261 + fig.legend(handles=[Patch(facecolor=fs.OI['grey'], alpha=0.25, label='Days with VIX close above 25')], loc='outside lower center')
262 + fs.save(fig, 'fig_rq3_series')
263 + n = len(d); clip0 = float((d['implied_corr_raw'] <= 0).mean() * 100) if 'implied_corr_raw' in d else np.nan
264 + share_hi = float(hi_vix.mean() * 100)
265 + cap = (rf"Implied and realized correlation among S\&P~500 constituents, 2010--2025 ({num(n)} trading days, series rebuilt from the shipped parquets by stage~1b of "
266 + rf"\texttt{{scripts/04}}, 29 constituents, days with fewer than 20 constituents dropped). Panel (a): the CBOE-style implied correlation from the ATM implied "
267 + rf"variances of SPX and of the constituents (eq.~(\ref{{eq:implcorr}}), clipped to $[0,1]$) and the 22-day pairwise realized correlation; panel (b): the log ratio "
268 + rf"that enters the stress regressions (mean {_txt(d['log_ratio'].mean(),3)}, standard deviation {_txt(d['log_ratio'].std(),3)}); panel (c): the VIX close from the "
269 + rf"complete daily series of the FinData OS lake. Shaded: days on which the VIX closes above 25 ({_txt(share_hi,1)}\% of days), the level component of the "
270 + rf"post-origin stress definition. Source: \texttt{{data/processed/correlation\_divergence\_v2.parquet}} (regenerable; diagnostics in Table~\ref{{tab:rq3_diag}}).")
271 + figure('fig_rq3_series', cap, 'fig:rq3_series')
272 +
273 +
274 +# ----------------------------------------------------------------------------
275 +# core: IRFs with bands
276 +# ----------------------------------------------------------------------------
277 +def fig_irf_bands():
278 + df = pd.read_csv(R / "irf_bands.csv")
279 + df = df[df['spec'] == 'levels']
280 + fig, axes = fs.new(width=6.5, height=2.9, nrows=1, ncols=2, sharex=True)
281 + panels = [('rv_to_ivshock', '(a) RV response to a one-SD IV shock'), ('iv_to_rvshock', '(b) IV response to a one-SD RV shock')]
282 + for ax, (key, title) in zip(axes, panels):
283 + for i, (ordering, label) in enumerate((('rvfirst', 'RV ordered first'), ('ivfirst', 'IV ordered first'))):
284 + sub = df[df['ordering'] == ordering].sort_values('horizon')
285 + ax.fill_between(sub['horizon'], sub[f'{key}_p05'], sub[f'{key}_p95'], color=fs.SERIES[i], alpha=0.18, lw=0, hatch=None if i == 0 else '///', edgecolor=fs.SERIES[i])
286 + ax.plot(sub['horizon'], sub[f'{key}_mean'], color=fs.SERIES[i], ls=fs.LINESTYLES[i], marker=fs.MARKERS[i], markevery=4, label=label)
287 + ax.axhline(0, color=fs.INK, lw=0.7)
288 + fs.panel_label(ax, title)
289 + ax.set_xlabel('Horizon (trading days)')
290 + ax.margins(x=0.02)
291 + axes[0].set_ylabel('Response (standard-deviation units)')
292 + axes[0].legend(loc='upper right')
293 + fs.save(fig, 'fig_irf_bands')
294 + n = int(df['n_tickers'].max())
295 + rv1 = df[(df['ordering'] == 'rvfirst')].sort_values('horizon')
296 + peak = rv1.loc[rv1['rv_to_ivshock_mean'].idxmax()]
297 + cap = (rf"Orthogonalized impulse responses of the bivariate VAR(5) in ATM implied volatility and realized variance (standardized levels), averaged across the "
298 + rf"{n} tickers with at least 1{{,}}000 complete observations. Lines: cross-ticker mean response to a one-standard-deviation shock under the two recursive "
299 + rf"orderings (solid, circles: RV ordered first, so that RV shocks move IV contemporaneously but not the reverse; dashed, squares: IV first); shaded bands "
300 + rf"(plain / hatched): cross-ticker average of the per-ticker residual-bootstrap 5th and 95th percentiles (500 draws, seed 42). "
301 + rf"Under the RV-first ordering the RV response to an IV shock peaks at {_txt(peak['rv_to_ivshock_mean'],3)} SD at horizon {int(peak['horizon'])} "
302 + rf"(band [{_txt(peak['rv_to_ivshock_p05'],3)}, {_txt(peak['rv_to_ivshock_p95'],3)}]). Variance decompositions and stationarity diagnostics are in "
303 + rf"Table~\ref{{tab:fevd_v2}}. Source: \texttt{{results/irf\_bands.csv}} (package L1-E8).")
304 + figure('fig_irf_bands', cap, 'fig:irf_bands')
305 +
306 +
307 +# ----------------------------------------------------------------------------
308 +# appendix: specification curve
309 +# ----------------------------------------------------------------------------
310 +def fig_spec_curve():
311 + d = pd.read_csv(R / "rq1_spec_curve.csv").sort_values('t_stat').reset_index(drop=True)
312 + d['x'] = np.arange(len(d))
313 + se_types = list(dict.fromkeys(d['se_type']))
314 + dims = [('se_type', 'Standard errors'), ('cutoff', 'Winsorization'), ('sample', 'Sample'), ('horizon_days', 'Horizon (days)'), ('controls', 'Controls'), ('return_basis', 'Return basis')]
315 + levels = [(k, lab, [str(v) for v in dict.fromkeys(d[k].astype(str))]) for k, lab, in dims]
316 + n_rows = sum(len(v) for _, _, v in levels)
317 + fig, (ax, axd) = fs.new(width=6.5, height=5.2, nrows=2, ncols=1, sharex=True, gridspec_kw={'height_ratios': [2.2, n_rows / 6]})
318 + for i, se in enumerate(se_types):
319 + s = d[d['se_type'] == se]
320 + ax.scatter(s['x'], s['t_stat'], s=9, marker=fs.MARKERS[i % 6], color=fs.SERIES[i % 6], edgecolor=fs.INK, linewidth=0.2, label=se.replace('_', ' '))
321 + for yv, lab in ((1.96, '$|t| = 1.96$'), (-1.96, None), (3.02, '$|t| = 3.02$ (Bonferroni threshold of version 1.2)'), (-3.02, None)):
322 + ax.axhline(yv, color=fs.INK2, lw=0.7, ls=':' if abs(yv) > 3 else '--')
323 + if lab:
324 + ax.text(len(d) * 0.01, yv + 0.1, lab, fontsize=7, color=fs.INK2, va='bottom')
325 + ax.axhline(0, color=fs.INK, lw=0.7)
326 + ax.set_ylabel('$t$-statistic of Skew$_{25\\delta}$')
327 + ax.legend(loc='upper left', ncol=2, fontsize=7)
328 + fs.panel_label(ax, f'(a) Sorted $t$-statistics over {len(d)} specifications')
329 + yt, ytl = [], []
330 + row = 0
331 + for k, lab, vals in levels:
332 + for v in vals:
333 + s = d[d[k].astype(str) == v]
334 + axd.scatter(s['x'], np.full(len(s), row), s=3, marker='|', color=fs.INK, linewidth=0.6)
335 + yt.append(row); ytl.append(f"{lab}: {v.replace('_', ' ')}" if v == vals[0] else f"{v.replace('_', ' ')}")
336 + row += 1
337 + axd.axhline(row - 0.5, color=fs.RULE, lw=0.5)
338 + axd.set_yticks(yt); axd.set_yticklabels(ytl, fontsize=6.5)
339 + axd.set_ylim(-0.6, row - 0.4); axd.invert_yaxis()
340 + axd.set_xlabel('Specifications ordered by $t$-statistic')
341 + axd.grid(False)
342 + fs.panel_label(axd, '(b) Specification indicators')
343 + fs.save(fig, 'fig_spec_curve')
344 + cap = (rf"Specification curve of the 25-delta skew coefficient in the return regressions ({len(d)} specifications: winsorization cutoff $\times$ standard-error "
345 + rf"estimator $\times$ sample $\times$ horizon $\times$ controls $\times$ return basis; Table~\ref{{tab:spec_curve}}). Panel (a): $t$-statistics sorted in increasing "
346 + rf"order, marker and colour by standard-error estimator; dashed lines at $\pm1.96$, dotted at $\pm3.02$. Panel (b): the specification behind each point. "
347 + rf"The median $t$ is {_txt(d['t_stat'].median())}; $|t|>1.96$ in {_txt(d['abs_t_gt_196'].mean()*100,1)}\% and $|t|>3.02$ in {_txt(d['abs_t_gt_302'].mean()*100,1)}\% "
348 + rf"of specifications; among the Driscoll--Kraay specifications the share with $|t|>3.02$ is "
349 + rf"{_txt(d[d['se_type'].str.startswith('DK')]['abs_t_gt_302'].mean()*100,1)}\%. Source: \texttt{{results/rq1\_spec\_curve.csv}} (package L1-C5).")
350 + figure('fig_spec_curve', cap, 'fig:spec_curve')
351 +
352 +
353 +# ----------------------------------------------------------------------------
354 +# appendix: price magnet
355 +# ----------------------------------------------------------------------------
356 +def fig_magnet():
357 + d = pd.read_csv(R / "rq4_price_magnet_v2.csv")
358 + fig, axes = fs.new(width=6.5, height=2.9, nrows=1, ncols=2)
359 + for ax, (kind, title, strip) in zip(axes, [('distance', '(a) By distance of the open to the strike', 'Distance '), ('oi_quintile', '(b) By open-interest concentration quintile', 'OI quintile ')]):
360 + s = d[d['kind'] == kind].reset_index(drop=True)
361 + x = np.arange(len(s))
362 + ax.errorbar(x - 0.12, s['null_share_mean'] * 100, yerr=[(s['null_share_mean'] - s['null_p025']) * 100, (s['null_p975'] - s['null_share_mean']) * 100],
363 + fmt=fs.MARKERS[1], color=fs.SERIES[1], capsize=3, lw=1, label='Permutation null (mean, 95% interval)', markeredgecolor=fs.INK, markeredgewidth=0.3)
364 + ax.errorbar(x + 0.12, s['observed_share'] * 100, yerr=1.96 * s['se_observed_date_clustered'] * 100,
365 + fmt=fs.MARKERS[0], color=fs.SERIES[0], capsize=3, lw=1, label='Observed share ($\\pm 1.96$ date-clustered SE)', markeredgecolor=fs.INK, markeredgewidth=0.3)
366 + ax.plot(x, s['mirror_placebo_share'] * 100, ls='', marker='x', color=fs.INK2, label='Mirror-strike placebo')
367 + ax.axhline(50, color=fs.INK, lw=0.7, ls=':')
368 + ax.set_xticks(x); ax.set_xticklabels([g.replace(strip, '').replace('Q1_Low', 'Q1 (low)').replace('Q5_High', 'Q5 (high)') for g in s['group']], rotation=0, fontsize=7.5)
369 + ax.set_ylabel('Days closing nearer the max-OI strike (%)')
370 + fs.panel_label(ax, title)
371 + ax.margins(y=0.2)
372 + for xx, (n, p) in enumerate(zip(s['n_obs'], s['p_value_perm'])):
373 + ax.annotate(f"n = {int(n):,}\np = {p:.2f}", xy=(xx, ax.get_ylim()[0]), xytext=(0, 2), textcoords='offset points', ha='center', va='bottom', fontsize=6.3, color=fs.INK2)
374 + axes[0].legend(loc='center right', fontsize=6.8)
375 + fs.save(fig, 'fig_magnet')
376 + a = d[d['kind'] == 'all'].iloc[0]
377 + cap = (rf"Price-magnet test on the shipped panel ({num(a['n_obs'])} ticker-days of eight stocks, 2010--2025, open within 0.1\%--10\% of the maximum-open-interest "
378 + rf"strike among options expiring within 30 days). Circles: observed share of days on which the close is relatively closer to that strike than the open, with "
379 + rf"$\pm1.96$ date-clustered standard errors; squares: the permutation null that re-pairs each (day, distance, strike side) with the open-to-close move of a random "
380 + rf"other day of the same ticker ({int(a['n_perm'])} draws, seed {int(a['seed'])}), mean and 95\% interval; crosses: the mirror-strike placebo. The dotted line is the "
381 + rf"50\% benchmark of version~1.2, which is not the correct null because a driftless move reaches a strike at distance $d$ only if it has the right sign and does not "
382 + rf"overshoot. Over all ticker-days the observed share is {_txt(a['observed_share']*100,2)}\% against a null of {_txt(a['null_share_mean']*100,2)}\% "
383 + rf"[{_txt(a['null_p025']*100,1)}, {_txt(a['null_p975']*100,1)}] (permutation $p$ = {a['p_value_perm']:.2f}; minimum detectable difference "
384 + rf"{_txt(a['mde_80pct_power_pp'],1)} pp at 80\% power). Source: \texttt{{results/rq4\_price\_magnet\_v2.csv}} (Table~\ref{{tab:magnet_v2}}).")
385 + figure('fig_magnet', cap, 'fig:magnet')
386 +
387 +
388 +# ----------------------------------------------------------------------------
389 +# appendix: RQ5 clipping dependence
390 +# ----------------------------------------------------------------------------
391 +def fig_rq5_clipping():
392 + c = pd.read_csv(R / "rq5_clipping_variant.csv")
393 + c = c[c['period'] != '2016-2025'].copy()
394 + c['year'] = c['period'].astype(int)
395 + rules = list(dict.fromkeys(c['rule']))
396 + rule_lab = {'features clipped at training quantiles': 'HAR-RV, regressors clipped at training 1%/99% quantiles',
397 + 'features unclipped': 'HAR-RV, regressors unclipped', 'log-RV model': 'log-RV HAR specification'}
398 + fig, ax = fs.new(width=6.5, height=3.0)
399 + years = sorted(c['year'].unique())
400 + for i, rule in enumerate(rules):
401 + s = c[(c['rule'] == rule) & (c['model'] == 'OLS: HAR-RV')].sort_values('year')
402 + ax.plot(s['year'], s['r2_oos'], color=fs.SERIES[i], marker=fs.MARKERS[i], ls=fs.LINESTYLES[i], label=rule_lab.get(rule, rule))
403 + v = c[(c['rule'] == rules[0]) & (c['model'] == 'OLS: HAR-RV')].sort_values('year')
404 + ax.plot(v['year'], v['r2_oos_vix'], color=fs.INK, marker='x', ls='-', lw=1.0, label='VIX benchmark, $(VIX_t/100)^2/252$')
405 + ax.axhline(0, color=fs.INK, lw=0.7)
406 + # axis truncated below the lowest HAR value so that the HAR lines stay readable; off-scale VIX values are marked
407 + floor = float(c[c['model'] == 'OLS: HAR-RV']['r2_oos'].min()) - 0.35
408 + ax.set_ylim(floor, max(float(c['r2_oos'].max()), float(v['r2_oos_vix'].max())) + 0.25)
409 + for _, rr in v[v['r2_oos_vix'] < floor].iterrows():
410 + ax.annotate(f"VIX {int(rr['year'])}: {rr['r2_oos_vix']:.2f} (off scale)", xy=(rr['year'], floor), xytext=(0, 14), textcoords='offset points',
411 + ha='center', va='bottom', fontsize=7, color=fs.INK2, arrowprops=dict(arrowstyle='-|>', color=fs.INK2, lw=0.7))
412 + ax.set_xticks(years)
413 + ax.set_xlabel('Test year (models re-estimated on 2010 to the previous year); vertical axis truncated below the lowest HAR-RV value')
414 + ax.set_ylabel('$R^2_{OOS}$, next-day SPX realized variance')
415 + fig.legend(loc='outside lower center', ncol=2, fontsize=7.5)
416 + fs.save(fig, 'fig_rq5_clipping')
417 + full = pd.read_csv(R / "rq5_clipping_variant.csv"); full = full[full['period'] == '2016-2025']
418 + g = lambda rule, m='OLS: HAR-RV': full[(full['rule'] == rule) & (full['model'] == m)].iloc[0]
419 + clip, unc = g(rules[0]), g('features unclipped')
420 + y20c = c[(c['rule'] == rules[0]) & (c['model'] == 'OLS: HAR-RV') & (c['year'] == 2020)]['r2_oos'].iloc[0]
421 + y20u = c[(c['rule'] == 'features unclipped') & (c['model'] == 'OLS: HAR-RV') & (c['year'] == 2020)]['r2_oos'].iloc[0]
422 + cap = (rf"Dependence of the RQ5 benchmark comparison on the treatment of the test-year regressors (SPX, next-day realized variance, rolling annual re-estimation "
423 + rf"2016--2025, FinData OS lake data, \texttt{{scripts/32}}). Lines: out-of-sample $R^2$ of the HAR-RV model by test year (vertical axis truncated below the lowest HAR-RV value; off-scale values of the VIX benchmark are marked with their value) when the test-year regressors are clipped at "
424 + rf"the training window's 1\%/99\% quantiles (the scheme of the RQ2 design), left unclipped, or when the model is estimated in logs; crosses: the VIX benchmark. "
425 + rf"Clipping caps $RV_t$ in 2020 far below its realized level: the HAR-RV $R^2_{{OOS}}$ for 2020 is {_txt(y20c,3)} clipped against {_txt(y20u,3)} unclipped, and over "
426 + rf"2016--2025 it is {_txt(clip['r2_oos'],3)} clipped against {_txt(unc['r2_oos'],3)} unclipped, versus {_txt(clip['r2_oos_vix'],3)} for the VIX benchmark "
427 + rf"(Diebold--Mariano $t$ of the unclipped HAR-RV against the VIX: {_txt(unc['dm_vs_vix_t'])}). Source: \texttt{{results/rq5\_clipping\_variant.csv}} (Table~\ref{{tab:rq5_v2}}, Panel C).")
428 + figure('fig_rq5_clipping', cap, 'fig:rq5_clipping')
429 +
430 +
431 +def main():
432 + print("=" * 70); print("MANUSCRIPT FIGURES (Stage E)"); print("=" * 70)
433 + config.ensure_output_dirs()
434 + for fn in [fig_coverage, fig_rq2_loss_ratios, fig_rq2_subperiod, fig_rq3_headline, fig_rq3_series, fig_irf_bands,
435 + fig_spec_curve, fig_magnet, fig_rq5_clipping]:
436 + print(fn.__name__)
437 + fn()
438 + print("\nFIGURES COMPLETE.")
439 +
440 +
441 +if __name__ == "__main__":
442 + main()
443
Menu

Theme
↑↓navigate ↵open escclose /⌘Kanywhere