NIPS26Repo's picture
Initial anonymized artifact release
149513d verified
|
Raw
History Blame Contribute Delete
4.4 kB

2026-03-23 CN 10 Tickers / 6 Months Multi-Persona Efficiency Ablation

Summary

This note consolidates three completed CN 10 tickers / 6m multi-persona live backtests into a reusable execution-efficiency ablation. The purpose is not to compare returns, but to quantify how much the optimized execution stack speeds up controlled long-horizon persona studies.

The three runs hold the business setting fixed:

  • mode: multi-personality
  • market: CN
  • tickers: 600519, 000858, 300750, 601318, 600036, 601888, 600276, 002415, 601899, 000333
  • backtest period: 2025-09-01 ~ 2026-02-28
  • trading days: 102
  • initial cash: ¥100,000
  • analysts: fundamental, technical, company_news
  • personas: conservative, balanced, aggressive, passive, equal_weight_index
  • benchmark mode: index
  • benchmark index: 000300.SH
  • LLM: enabled

Experimental Conditions

E0: Baseline

  • comparison run id: 20260308_032534_600248
  • report: reports/multi_personality/20260308_032534_600248/comparison_report.md
  • data: reports/multi_personality/20260308_032534_600248/comparison_data.json

E1: Optimized-Cold

  • comparison run id: 20260308_164113_268847
  • log: logs/cn10_mp_20260308_164108.log
  • report: reports/multi_personality/20260308_164113_268847/comparison_report.md
  • data: reports/multi_personality/20260308_164113_268847/comparison_data.json

E2: Optimized-Warm

  • comparison run id: 20260309_120759_016922
  • log: logs/cn10_mp_rerun_20260309_120714.log
  • report: reports/multi_personality/20260309_120759_016922/comparison_report.md
  • data: reports/multi_personality/20260309_120759_016922/comparison_data.json

Core Results

End-to-End Runtime

Experiment Run ID Total Duration (sec) Total Duration (hours) Relative to E0
E0 baseline 20260308_032534_600248 157429.02 43.73 1.00x
E1 optimized-cold 20260308_164113_268847 38828.22 10.79 4.05x
E2 optimized-warm 20260309_120759_016922 38965.22 10.82 4.04x

LLM Usage (comparison report accounting)

Experiment LLM Calls Total Tokens Estimated Cost
E0 baseline 11219 10255794 ¥12.0572
E1 optimized-cold 8160 9299500 ¥10.7245
E2 optimized-warm 8160 9264269 ¥10.6699

Shared Phase-One Statistics

Metric E0 E1 E2
artifact cache hits n/a 0 102
artifact cache misses n/a 102 0
sync load seconds n/a 82.08 0.19
prefetch submitted n/a 101 101
prefetch hits n/a 101 101
prefetch compute seconds n/a 7633.76 65.83
pipeline hidden seconds n/a 7633.76 65.82
pipeline utilization n/a 100.0% 100.0%
phase2 execution mode n/a parallel_threads parallel_threads

Interpretation

E0 -> E1: Execution-Model Gain

  • total duration: 157429.02s -> 38828.22s
  • absolute savings: 118600.80s (32.94 hours)
  • relative speedup: 4.05x
  • runtime reduction: 75.34%

In the same accounting convention:

  • LLM calls: 11219 -> 8160 (-27.27%)
  • total tokens: 10255794 -> 9299500 (-9.32%)
  • estimated cost: ¥12.0572 -> ¥10.7245 (-11.05%)

E1 -> E2: Warm-Cache Gain

  • total duration: 38828.22s -> 38965.22s
  • end-to-end change: +137.00s (effectively flat)

But at the phase-one level:

  • sync load: 82.08s -> 0.19s (-99.76%)
  • prefetch compute: 7633.76s -> 65.83s (-99.14%)
  • hidden pipeline seconds: 7633.76s -> 65.82s (-99.14%)

This shows that artifact caching nearly eliminates repeated phase-one work, while the main end-to-end bottleneck after E1 has already shifted to later persona-specific stages.

Notes on Accounting

The call/token/cost reductions should be interpreted as values under the comparison-report accounting scheme. In the older baseline run, shared phase-one token usage is counted separately in the comparison data, while the newer execution path reports shared phase-one activity differently. We therefore use these values as report-level observations rather than as a claim about exactly identical physical API savings under every instrumentation scheme.