"""Generate paper/results_macros.tex from saved machine-readable outputs. Every numerical claim in the manuscript resolves to a macro defined here (CLAUDE.md §26.5). Rerun after any estimation stage; compile refuses stale numbers by construction because macros are the only path into the text. """ from __future__ import annotations import json from pathlib import Path import polars as pl ROOT = Path(__file__).resolve().parents[1] OUTE = ROOT / "outputs" M = [] def macro(name: str, value: str) -> None: M.append(f"\\newcommand{{\\{name}}}{{{value}}}") def num(x, nd=0): return f"{x:,.{nd}f}" def main() -> None: # ---- data scale (stage C summary) sc = json.loads((OUTE / "diagnostics/stage_c/stage_c_summary.json").read_text()) macro("DataRows", num(sc["total_rows"])) macro("DataRowsBn", f"{sc['total_rows']/1e9:.2f}") macro("DataTickers", num(sc["tickers_total"])) macro("DataSpells", num(sc["spells_total"])) macro("DataOHLCViol", num(sc["total_ohlc_violations"])) macro("DataDupKeys", num(sc["total_duplicate_keys"])) macro("DataMultiSpellTickers", num(sc["multi_spell_tickers"])) # ---- splits (stage D) ev = pl.read_parquet(OUTE / "diagnostics/stage_d/split_events.parquet") fwd = int(ev["event"].str.starts_with("split").sum()) macro("NSplitEvents", num(ev.height)) macro("NSplitFwd", num(fwd)) macro("NSplitRev", num(ev.height - fwd)) # ---- extreme minute returns (fig E5 source) e5 = pl.read_csv(OUTE / "figures/data/figE5_extreme_minutes.csv") g1_90s = e5.filter(pl.col("year").is_between(1992, 1999))["gt1"].mean() * 1e4 g1_10s = e5.filter(pl.col("year").is_between(2010, 2019))["gt1"].mean() * 1e4 g5_1719 = e5.filter(pl.col("year").is_between(2017, 2019))["gt5"].mean() * 1e4 g5_20 = e5.filter(pl.col("year") == 2020)["gt5"].mean() * 1e4 macro("ExtremeGtOneNineties", num(g1_90s, 0)) macro("ExtremeGtOneTens", num(g1_10s, 1)) macro("ExtremeDropFactor", num(g1_90s / g1_10s, 0)) macro("ExtremeGtFivePreCovid", f"{g5_1719:.3f}") macro("ExtremeGtFiveCovid", num(g5_20, 1)) macro("ExtremeCovidFactor", num(g5_20 / g5_1719, 0)) # ---- Epps (fig E8 source) ep = pl.read_csv(OUTE / "diagnostics/stage_e/epps_curves.csv") def epv(y, k): return ep.filter((pl.col("year") == y) & (pl.col("agg_minutes") == k) )["mean_pairwise_corr"][0] macro("EppsNinetySixOne", f"{epv(1996,1):.3f}") macro("EppsNinetySixThirty", f"{epv(1996,30):.2f}") macro("EppsFifteenOne", f"{epv(2015,1):.2f}") macro("EppsFifteenThirty", f"{epv(2015,30):.2f}") macro("EppsTwentyFourOne", f"{epv(2024,1):.2f}") macro("EppsTwentyFourThirty", f"{epv(2024,30):.2f}") # ---- stage F baseline f1 = json.loads((OUTE / "estimates/stage_f_baseline.json").read_text()) macro("NEventsReal", num(f1["n_events_real"])) macro("NEventsRealMM", f"{f1['n_events_real']/1e6:.2f}") macro("NEventsPlacebo", num(f1["n_events_placebo"])) for r in f1["baseline"]: if r["h"] in (5, 30): tag = "Five" if r["h"] == 5 else "Thirty" macro(f"DDBreadth{tag}", f"{r['dd_breadth']*1e4:.2f}") macro(f"DDBreadth{tag}SE", f"{r['se_breadth']*1e4:.2f}") grad = {g["group"]: g for g in f1["gradient_state"]} macro("DDGradLargest", f"{grad['12–20']['dd']*1e4:.2f}") macro("DDGradLargestSE", f"{grad['12–20']['se']*1e4:.2f}") # ---- stage F2 (if available) f2p = OUTE / "estimates/stage_f2.json" if f2p.exists(): f2 = json.loads(f2p.read_text()) nw = {(r["scope"], r["h"]): r for r in f2["network"]} for h, tag in [(5, "Five"), (15, "Fifteen"), (30, "Thirty")]: macro(f"NetDiff{tag}", f"{nw[('all',h)]['dd']*1e4:.1f}") macro(f"NetDiff{tag}SE", f"{nw[('all',h)]['se']*1e4:.1f}") macro("NetDiffFiveNineties", f"{nw[('1992–1999',5)]['dd']*1e4:.0f}") macro("NetDiffFiveOughts", f"{nw[('2000–2009',5)]['dd']*1e4:.0f}") macro("NetDiffFiveTens", f"{nw[('2010–2019',5)]['dd']*1e4:.0f}") macro("NetDiffFiveTwenties", f"{nw[('2020–2026',5)]['dd']*1e4:.0f}") cm = {(r["scope"], r["h"]): r for r in f2["cumulative"]} macro("CumThirty", f"{cm[('all',30)]['dd']*1e4:.1f}") macro("CumThirtySE", f"{cm[('all',30)]['se']*1e4:.1f}") cj = pl.DataFrame(f2["cojumps_yearly"]) def sev(y): return cj.filter(pl.col("year") == y)["mean_share4"][0] * 100 macro("CoJumpSevCalm", f"{cj.filter(pl.col('year').is_in([2015,2017,2019]))['mean_share4'].mean()*100:.1f}") macro("CoJumpSevGFC", f"{sev(2008):.1f}") macro("CoJumpSevCovid", f"{sev(2020):.1f}") sv = cj.filter(pl.col("year") >= 2013) macro("SevereCovid", f"{cj.filter(pl.col('year')==2020)['severe_per1000'][0]:.1f}") macro("SeverePreCovid", f"{cj.filter(pl.col('year').is_in([2017,2018,2019]))['severe_per1000'].mean():.2f}") macro("SevereGFC", f"{cj.filter(pl.col('year')==2008)['severe_per1000'][0]:.1f}") macro("SevereEarlyNineties", f"{cj.filter(pl.col('year').is_between(1992,1997))['severe_per1000'].mean():.0f}") macro("CoJumpSevEarlyNineties", f"{cj.filter(pl.col('year').is_between(1992,1997))['mean_share4'].mean()*100:.1f}") macro("CoJumpPerThousandRecent", f"{cj.filter(pl.col('year').is_between(2021,2025))['per1000'].mean():.0f}") # ---- zero-return share (fig E7 source) and post-decimalization extremes e7 = pl.read_csv(OUTE / "figures/data/figE7_zeroret.csv") macro("ZeroRetNinetySix", f"{e7.filter(pl.col('year')==1996)['zero_pct'][0]:.0f}") macro("ZeroRetTwentyFour", f"{e7.filter(pl.col('year')==2024)['zero_pct'][0]:.0f}") macro("ExtremeGtOneTwoThousandTwo", num(e5.filter(pl.col("year") == 2002)["gt1"][0] * 1e4, 0)) macro("ExtremeGtOneNinetyTwo", num(e5.filter(pl.col("year") == 1992)["gt1"][0] * 1e4, 0)) # ---- revision checks (scripts/17_revision_checks.py, 2026-10-17) rcp = OUTE / "estimates/revision_checks.json" if rcp.exists(): rc = json.loads(rcp.read_text()) pb = rc["pre_event"]["breadth"] macro("PreBreadthReal", num(pb["real"] * 1e4, 0)) macro("PreBreadthPlacebo", num(pb["placebo"] * 1e4, 0)) macro("PreBreadthDiff", f"{pb['diff']*1e4:.1f}") macro("PreBreadthDiffSE", f"{pb['se']*1e4:.1f}") pi_ = rc["pre_event"]["intensity"] macro("PreIntensityDiff", f"{pi_['diff']:.3f}") macro("PreIntensityDiffSE", f"{pi_['se']:.3f}") br = rc["base_rates"] macro("BaseBreadthShare", f"{br['placebo_breadth_5']:.2f}") macro("NEligMean", num(br["nelig_5_mean"], 0)) macro("NEligNineties", num(br["nelig_5_by_decade"]["1992–1999"], 0)) macro("EventsPerStockWeek", f"{br['events_per_stock_week']:.1f}") macro("EventDays", num(br["event_days"])) macro("NeighborBaseShare", f"{br['placebo_nbh_5']:.2f}") macro("MatchRate", f"{br['match_rate']*100:.1f}") macro("MatchRateNineties", f"{br['match_rate_by_decade']['1992–1999']*100:.0f}") macro("MatchRateMinPost", f"{min(v for k, v in br['match_rate_by_decade'].items() if k != '1992–1999')*100:.0f}") macro("CumBaseFifteen", f"{br['placebo_cum_15']*100:.0f}") mde = 2.8 * f1["baseline"][2]["se_breadth"] # h=5, 5% size / 80% power macro("MDEFive", f"{mde*1e4:.1f}") macro("MDEEventsPer", num(1.0 / (mde * br["nelig_5_mean"]), 0)) ca = rc["coarrival"] macro("ShareSingleton", f"{ca['share_by_class']['1']*100:.0f}") macro("MeanCoArrival", f"{ca['mean_events_per_event_minute']:.1f}") cls = {r["cls"]: r for r in ca["by_class"]} macro("DDSingletonFive", f"{cls['1']['db5']*1e4:.1f}") macro("DDSingletonFiveSE", f"{cls['1']['db5_se']*1e4:.1f}") macro("DDSingletonThirty", f"{cls['1']['db30']*1e4:.1f}") macro("DDSingletonThirtySE", f"{cls['1']['db30_se']*1e4:.1f}") macro("DDMultiThirty", f"{cls['3–5']['db30']*1e4:.1f}") macro("DDMultiThirtySE", f"{cls['3–5']['db30_se']*1e4:.1f}") macro("NetDiffSingleton", f"{cls['1']['nd5']*1e4:.0f}") macro("NetDiffSingletonSE", f"{cls['1']['nd5_se']*1e4:.1f}") st = {r["h"]: r for r in rc["stratified"]} macro("PostDiffFive", f"{st[5]['post_diff']*1e4:.0f}") macro("StratFive", f"{st[5]['strat']*1e4:.0f}") macro("StratFiveSE", f"{st[5]['strat_se']*1e4:.1f}") macro("StratThirty", f"{st[30]['strat']*1e4:.0f}") macro("StratThirtySE", f"{st[30]['strat_se']*1e4:.1f}") macro("StratShareOfBase", f"{st[5]['strat']/br['placebo_breadth_5']*100:.0f}") macro("NetDiffFiveTickerSE", f"{rc['cluster']['nd5']['se_ticker']*1e4:.1f}") bs = {r["sign"]: r for r in rc["network_by_sign"]} macro("NetDiffNeg", f"{bs['negative']['dd']*1e4:.0f}") macro("NetDiffPos", f"{bs['positive']['dd']*1e4:.0f}") nc = {r["scope"]: r for r in rc["network_components"]} macro("NetRealNdNineties", f"{nc['1992–1999']['real_nd_5']*1e4:.0f}") macro("NetPlaceboNdNineties", f"{nc['1992–1999']['placebo_nd_5']*1e4:.0f}") macro("NetRealNdTens", f"{nc['2010–2019']['real_nd_5']*1e4:.0f}") macro("NetPlaceboNdTens", f"{nc['2010–2019']['placebo_nd_5']*1e4:.0f}") tex = "% Auto-generated by scripts/15_build_results_macros.py — DO NOT EDIT.\n" tex += "\n".join(M) + "\n" (ROOT / "paper" / "results_macros.tex").write_text(tex) print(f"{len(M)} macros written") if __name__ == "__main__": main()