ai-asset-pricing: Skill for Claude Code

/run — Fast Portfolio Sort Execution

Executes batch or single portfolio sorts on the Dickerson bond dataset without spawning an agent. One Bash call, results saved automatically.

Arguments

/run <mode> <signals...> [options]

Argument	Values	Default
mode	batch, single	required
signals	space-separated signal names (e.g., cs ytm bbtm)	required
--wfs	Use WithinFirmSort instead of SingleSort	off
--rating	IG, NIG, or omit for all	all
--nport	Number of portfolios (SingleSort only)	5
--hp	Holding period	1

Execution Steps

1. Parse arguments from the user's /run command
2. Determine the bond data path — look for Dickerson bond parquet in data/ (check metadata.json for datasets from bonds-wrds-expert), or use the path from canonical local state reported by tools/bootstrap.py audit
3. Generate the Python script from the template below (fill in signals, options, data path)
4. Execute via single Bash call using the Python path from canonical local state reported by tools/bootstrap.py audit
5. Print the mandatory summary table:

| Signal | Weight | Mean(%) | t-stat | SR | Turn(%) |

All values annualized (%). Turnover = avg monthly two-way (%).

1. Follow with 2-3 sentences interpreting the results. Always state the sample period (first and last date in the returns).

Date Filtering

The TRACE-aligned sample filter (2002-08 to 2024-12) should only be applied when:

• The data is corporate bond / bond data, AND
• The user explicitly requests a TRACE-aligned sample

Otherwise, use the full date range in the data as-is. Do NOT hardcode the TRACE filter by default.

Template: Batch SingleSort

import warnings; warnings.filterwarnings('ignore')
import pandas as pd
import numpy as np
import PyBondLab as pbl
from PyBondLab.report import ResultsReporter
from PyBondLab.describe.utils import compute_nw_tstat

SIGNALS = {signals}          # e.g., ['cs', 'ytm', 'bbtm']
RATING = {rating}            # None, 'IG', or 'NIG'
NPORT = {nport}              # e.g., 5
HP = {hp}                    # e.g., 1
MNEMONIC = '{mnemonic}'      # e.g., 'batch_3s'

SCRIPT = '<generated by /run skill>'

data = pd.read_parquet('{data_path}')
# TRACE-aligned filter: ONLY apply if corporate bond data AND user requests TRACE-aligned sample
# data = data[(data['date'] >= '2002-08-01') & (data['date'] <= '2024-12-31')]
data['spc_rat'] = data['spc_rat'].astype('float64')

batch = pbl.BatchStrategyFormation(
    data=data, signals=SIGNALS,
    holding_period=HP, num_portfolios=NPORT,
    turnover=True, rating=RATING,
    columns={'ID': 'cusip', 'ret': 'ret_vw', 'VW': 'mcap_e', 'RATING_NUM': 'spc_rat'},
    n_jobs=-2, verbose=False,
)
results = batch.fit()

report_path = ResultsReporter(results, mnemonic=MNEMONIC, script_text=SCRIPT).generate()

# Print summary table
rows = []
for sig in SIGNALS:
    r = results[sig]
    ew_ls, vw_ls = r.get_long_short()
    for label, ls in [('EW', ew_ls), ('VW', vw_ls)]:
        valid = ls.dropna()
        ann_mean = valid.mean() * 12 * 100
        sr = valid.mean() / valid.std() * np.sqrt(12) if valid.std() > 0 else 0
        nw_lag = int(len(valid) ** 0.25)
        t = compute_nw_tstat(valid, nw_lag=nw_lag)[0]
        turn = '—'
        try:
            ew_t, vw_t = r.get_turnover()
            td = ew_t if label == 'EW' else vw_t
            if td is not None and not td.empty:
                turn = f'{td.mean().mean() * 100:.1f}'
        except: pass
        rows.append((sig, label, ann_mean, t, sr, turn))

print('|Signal|Weight|Mean(%)|t-stat|SR|Turn(%)|')
print('|---|---|---|---|---|---|')
for sig, wt, m, t, sr, turn in rows:
    print(f'|{sig}|{wt}|{m:.2f}|{t:.2f}|{sr:.3f}|{turn}|')
first_sig = SIGNALS[0]
r0 = results[first_sig]
ew0, _ = r0.get_long_short()
print(f'Sample: {ew0.dropna().index[0].strftime("%Y-%m")} to {ew0.dropna().index[-1].strftime("%Y-%m")}')
print(f'Report: {report_path}')