Datasets:

NIPS26Repo
/

quantarena-artifacts

metrics.json — summary metrics (total_return, max_drawdown, sharpe_ratio, total_trades, avg_position_days, avg_cash_ratio, avg_gross_exposure, value_filter_pass_rate, etc.)
trades.csv — per-trade log (date, ticker, action, shares, price, value, justification)
equity_curve.csv — daily portfolio time series (date, total_value, daily_return, cashflow=cash_balance, benchmark_value, benchmark_return)
backtest_report.md — human-readable summary

Multi-mandate comparison bundles live under <repo_root>/reports/multi_personality/<bundle>/ and contain:

comparison_data.json — mapping from mandate key to its run ID, return, MDD, trade count, and aggregated metrics
comparison_report.md — human-readable summary
personality_summary.csv — flat CSV summary

Experiment 1 — Six-month main case study (the headline result)

Mandate	US run (Sep 1 2025 – Feb 28 2026, 124 trading days)	CN run (Sep 1 2025 – Feb 28 2026, 102 trading days)
Fundamental Value	`mp_fundamental_value_20260409_165131_876468`	`mp_fundamental_value_20260413_163905_878961`
Macro Tactical	`mp_macro_tactical_20260409_165131_876505`	`mp_macro_tactical_20260413_163905_878979`
Behavioral Momentum	`mp_behavioral_momentum_20260409_165131_876499`	`mp_behavioral_momentum_20260413_163905_878974`
Low-Volatility (Smart Beta)	`mp_smart_beta_passive_20260409_165131_876510`	`mp_smart_beta_passive_20260413_163905_878984`
Equal-Weight (Baseline)	`mp_equal_weight_index_20260409_165131_876514`	`mp_equal_weight_index_20260413_163905_878987`

US comparison bundle: multi_personality/20260409_165121_176781/
CN comparison bundle: multi_personality/20260413_163854_715331/
Initial capital: $100,000 (each market)
Universe: 20 tickers per market on a 5×4 sector/style grid (Appendix Table~\ref{tab:sector_style_universe})
Backend: DeepSeek-V3.2

Experiment 2 — Full-run reproducibility (US 6M, Run 2)

Independent re-run of all five US 6M mandates with identical configuration to Experiment 1.

Mandate	US Run 2
Fundamental Value	`mp_fundamental_value_20260425_035029_413185`
Macro Tactical	`mp_macro_tactical_20260425_035029_413208`
Behavioral Momentum	`mp_behavioral_momentum_20260425_035029_413202`
Low-Volatility (Smart Beta)	`mp_smart_beta_passive_20260425_035029_413212`
Equal-Weight (Baseline)	`mp_equal_weight_index_20260425_035029_413216`

Experiment 3 — US 3M mechanism ablation (Dec 1 2025 – Feb 28 2026)

Each LLM-conditioned mandate has a Full variant (full doctrine) and one Ablated variant (a single guardrail removed).

Mandate / variant	Run ID
Fundamental Value — Full	`20260417_160942`
Fundamental Value — No filter	`20260417_191443`
Behavioral Momentum — Full	`20260417_220721`
Behavioral Momentum — No guardrails	`20260418_004927`
Macro Tactical — Full	`20260418_033940`
Macro Tactical — No tilt	`20260418_072633`
Low-Volatility — Reference	`20260418_110947`
Equal-Weight — Reference	`20260418_111210`

Experiment 4 — US 3M backend robustness (GPT-5.4 replacement)

Same 3M window as Experiment 3, but the reasoning backend swapped from DeepSeek-V3.2 to GPT-5.4 while every other component is held fixed.

Mandate	GPT-5.4 run
Fundamental Value	`20260423_124148`
Macro Tactical	`20260424_094955`
Behavioral Momentum	`20260424_132728_372104`
Low-Volatility	`20260424_094956`
Equal-Weight	`20260424_132723_271683`

These per-run metrics are also aggregated into derived/gpt54_robustness/backend_comparison.csv, which is what the generate_backend_robustness_figure.py script consumes.

Experiment 5 — Execution efficiency ablation (CN, 10 tickers, 6M)

Three execution paths (E0/E1/E2) on a fixed business setting (10 CN tickers, multi-mandate, Sep 1 2025 – Feb 28 2026), used only to demonstrate runtime/cost reductions enabled by shared phase-one preprocessing and warm artifact caching.

Path	Run ID	Wall-clock
E0 (baseline)	`20260308_032534_600248`	43.73 h
E1 (optimized cold-cache)	`20260308_164113_268847`	10.79 h
E2 (optimized warm-cache)	`20260309_120759_016922`	10.82 h

Source doc: docs/2026-03-23-cn-10tickers-6m-efficiency-ablation.md.