When Is Fresh AI Stock Analysis Worth It? Allocating a Limited Research Budget Across Price Updates, Deep Forecast Refreshes, and Editorial Reviews of Top Stock Picks
Description
When is another AI research run worth more than keeping that money invested? For long-horizon investment platforms covering hundreds of equities, fresh reasoning must improve decisions enough to cover its recurring cost. This resource-allocation problem underpins iPulse AI. Using an archive of 8,120 production reports on a balanced panel of 203 stocks across four production batches (retaining ten matched agent configurations across seven investment frameworks), we measure how much decision value can be retained by updating current prices against existing forecasts before paying for fresh multi-agent inference. Generating these reports consumed nearly 100 million output and reasoning tokens. Across three refresh transitions, we compare (1) old price targets repriced at current market quotes (with and without a volatility-scaled shock safeguard), (2) a zero-inference mechanical reset that re-anchors old predicted growth rates to current prices, and (3) fresh deep reports. In our primary economic evaluation--selecting equal-weight Top 10 USD portfolios from one-year forecasts net of 10 bps transaction costs and a USD 3,000 all-in research charge on a USD 300,000 account--shock-guarded old targets beat SPY across all three short (3-month) holding windows without incurring new inference costs, averaging 2.25% versus 1.66% for SPY and 4.24% for fresh reports in unchanged-model runs. The thesis half-life question becomes visible in the longer window: extending the first holding window to nearly six months without a refresh causes the guarded reuse portfolio to fall to -11.26%, whereas fresh reports return +8.55% (SPY: 3.46%). Supporting 90-day forecast diagnostics reveal why inexpensive baselines compete so effectively over shorter windows: mechanically resetting old forecast paths to current prices eliminates 98% of the Mean Absolute Error (MAE) gap between stale targets (15.26 pp) and fresh LLM analysis (13.51 pp vs. 13.55 pp for reset), yet still turns over 60%--90% of the Top 10 shortlist. Expanding the shortlist to Top 30 reverses the unchanged-model comparison in favor of guarded reuse (4.17% vs. 1.37% for fresh), and leave-one-out stress tests highlight strong dependence on individual outlier winners (MRNA). A separate audit of nine production editions shows substantial changes in published Top 10 membership between editions, already present in their incoming candidate lists. Within each edition, the final editorial screen retained its incoming list without replacements. Finally, a USD 300,000 account funding USD 3,000 full-universe runs shows why frequency matters: monthly budgets approach the realized short-window value of refreshes, while weekly budgets require substantial additional gains.
Version 2.1.0 integrates an additional ranking-policy experiment into Section 5.4 without changing the paper's premise, abstract, operating-cost assumptions or earlier findings. On the same 203-stock panel and costs, reuse 75/25 improves the two unchanged-model short-window Top 10 mean from 4.96% to 6.19% (1.23 percentage points); the best-sleeve cash control retains a 0.67-point benefit. Moderna dependence, Top 30 and the longer holding window limit a universal claim. Fresh research can still justify its cost as a thesis ages. The package adds a network-free adapter for 609 councils, 792 fee/cost scenarios, rank agreement and concentration arithmetic.
The USD300,000 account, USD3,000 full-cycle research bill, net dividends, fees and frozen cutoffs are unchanged. The 96,830,247 generated-token audit is preserved. Paper/derived exports CC BY4.0; focused code MIT; licensed raw market inputs remain excluded.
arXiv: https://arxiv.org/abs/2611.04348
GitHub: https://github.com/TheFutureEdge/research-when-fresh-ai-stock-analysis-is-worth-it/releases/tag/v2.1.0
Files
When_Is_Fresh_AI_Stock_Analysis_Worth_It_v2_1_0.pdf
Additional details
Related works
- Is identical to
- Preprint: arXiv:2611.04348 (arXiv)