发表机构
ACS Athens（雅典美国学院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究提出将共形预测区间与分数凯利结合用于投资组合头寸规模确定，经6年测试表现优于被动基准，还引入风险控制优化回撤，相关配置经预注册和LLM智能体搜索验证。
AI 中文摘要
共形预测传统上用于量化预测不确定性，我们将这种不确定性用于第二种用途，结合75%的共形区间与分数凯利来确定投资组合头寸规模：区间扩大时我们缩小头寸，区间缩小时我们扩大头寸。在6年的开发窗口（2016-2021年）中，考虑交易成本和严格的杠杆上限，该策略的年化净对数增长率复合为28.5%，夏普比率为1.34，最大回撤为27.7%，而持有标普500的年化增长率为15.9%，相同杠杆下被动投资组合的年化增长率为21-22%。我们在开发窗口的主要发现与时间序列共形预测的文献建议相悖：每一项使区间更快适应市场状况的调整都会使年化增长率损失0.7至5.3个百分点，最优方法是最简单的方法：缓慢、未加权、按资产的滚动分位数。当区间用于确定头寸规模而非描述单个预测时，宽度稳定性优于局部尖锐度，在相同杠杆下，它比教科书标准偏差的表现高出2.1个百分点。我们还实施了一项风险控制：当区间下行遗漏的次数远超过其历史概率时，我们降低杠杆。在开发窗口中，这将最大回撤从27.7%降至20.3%，同时提高了夏普比率，击败了所有40个安慰剂时间点（基于排名的p值为1/41）。这些数值来自自主LLM智能体对200种配置的搜索，因此我们将2022年及以后的数据封存，并在一次评估前预先注册了配置、基准和解释规则。校准效果良好（覆盖率为0.745，对应目标0.750，在2022年表现最弱）；增长率未达标：两种配置的年化收益率分别为8.5%和7.0%，低于被动基准，而预先注册的事后基准在原始增长率上击败了它们，同时承受了46%的回撤。所有结果均按预先注册的要求报告。
英文摘要
We size portfolio positions from conformal prediction intervals, reading the interval half-width as the scale in a capped fractional-Kelly rule (Conformal Kelly), and evaluate the rule under a pre-registered protocol. The development record is an autonomous LLM-agent search over roughly 200 configurations on a single six-year window (2016-2021) of eight US-listed ETFs. All later data was sealed before the search, and two final configurations, nine comparison books and the mapping from outcomes to conclusions were registered before unsealing. On the sealed 2022-2024 window, realized coverage was 0.745 against a 0.750 target, but the two configurations' annualised net log growth was 8.5% and 7.0% (28.5% and 25.8% in development), below an unlevered equal-weight book (9.5%), the leveraged passive books (15.6-16.8%) and a registered hindsight four-equity book (10.8%, at a 46% drawdown); both ranked last of eleven entries on Sharpe and Calmar. A post-hoc split, outside the registered protocol, shows the passing coverage averaging 0.592 in 2022 and 0.842 afterwards, with the entire loss in the under-covered year. The split does not show that under-coverage caused the loss: the leverage cap bound on every 2022 trading day, so narrow intervals could not enlarge the book, and a long-equity book in an equity bear market is the simpler account. But the registered acceptance test passed a sixteen-point conditional failure, which motivates, without establishing, conditional coverage as an endpoint for rules that read interval width as a scale. On the development window, a slow per-asset conformal quantile out-grew every locally adaptive alternative tested, but only under a binding leverage cap, and block-bootstrap intervals include zero for 11 of the 13 contrasts we could re-derive.
Comments31 pages, 9 figures. Extended version of a paper to appear at the 7th ACM International Conference on AI in Finance (ICAIF '26). v2: retitled and reorganized around the pre-registered out-of-sample result; adds dependence-aware block-bootstrap intervals and a post-hoc conditional-coverage diagnostic
