AI 证券与资管全景图 · 给资产定价的竞技场 · 2026-07 · 供给侧Securities & Asset Management × AI · The Pricing Arena · Jul 2026 · Supply side
客户的游艇在哪里?
Where are the customers’ yachts?
整个产业建立在一个不对称之上:费的基点旱涝保收,alpha 的基点镜花水月,产业的利润是两者之差。AI 在这个结构里的位置非常明确:它降低了「产生费用」的成本——研报、客服、合规、代码;却没有任何证据显示它提高了「产生 alpha」的能力。1940 年那个访客问「客户的游艇在哪里」,八十六年过去,唯一的变化是:游艇的燃料从佣金换成了管理费,现在正在换成算力。
The whole industry rests on one asymmetry: the fee’s basis points are guaranteed, alpha’s basis points are a mirage — the industry’s profit is the difference. AI’s place in this structure is precise: it cuts the cost of producing fees (research, service, compliance, code) while showing no evidence of improving the production of alpha. Eighty-six years after the visitor asked where the customers’ yachts were, the only change is the yacht’s fuel: commissions became management fees, and are now becoming compute.
本图相信:替代已在发生——高盛 CEO 称 AI 可在几分钟内起草 95% 的 IPO 招股书(传统需 6 人 2 周);摩根士丹利 98% 的顾问团队活跃使用 AI;中金财富投研文档效率 +90%;高盛/摩根大通/花旗被报道把初级分析师招聘班级最多削减三分之二(B)。This map believes: replacement is under way — Goldman’s CEO says AI drafts 95% of an IPO prospectus in minutes (vs 6 people × 2 weeks); 98% of Morgan Stanley advisor teams actively use AI; CICC Wealth reports +90% research-document throughput; Goldman, JPMorgan and Citi reportedly cut junior-analyst classes by up to two-thirds (B).
本图不相信:「AI 产生可持续 alpha」。AI 对冲基金指数 15 年年化 9.8%,同期标普 500 为 13.7%——赢了同行,输给指数;顶部四分位主动基金 4 年后留存率 0.0%;流传最广的「GPT 战胜分析师」论文已被撤回,且有结论相反的研究(A/B)。AI 是 alpha 的加速器,但加速的是它的衰减。It does not believe «AI generates sustainable alpha». The AI hedge-fund index annualised 9.8% over 15 years against the S&P 500’s 13.7% — beating peers, losing to the index; top-quartile active funds hold their quartile after 4 years at a rate of 0.0%; the most-quoted «GPT beats analysts» paper is withdrawn, with opposite findings published (A/B). AI accelerates alpha — its decay, not its creation.
主脊:一笔委托资金的八站旅程(获客→产品创设→投研→交易执行→托管清算→排名与费率→赎回轮回),①③标红;三条结构带:三层诚实账、监管三分岔、灰产前线;判断层:费率 > 持有行为 > 配置 > 选基 > 选股——排序最前的杠杆,恰是产业最不愿谈的两件事。三图分工:个人的钱怎么走→wealth,信用扩张→credit,套利与微观结构→trade,抽水赌局→bet,荐股黑嘴→scam。
The spine: one mandate’s eight tollbooths (acquisition → product → research → execution → custody → rankings & fees → redemption), ① and ③ in red; three bands: the honesty ledger, the regulatory fork, the grey frontier; the judgment layer: fees > holding behaviour > allocation > fund-picking > stock-picking — the top two levers are precisely what the industry least discusses. Sisters: wealth, credit, trade, bet, scam.
89.50%
15 年期跑输标普 500 的美国大盘主动基金比例;全部 22 个细分类别,无一类别多数主动基金跑赢基准(SPIVA,A)Share of US large-cap active funds losing to the S&P 500 over 15 years; across all 22 categories, in none does a majority beat the benchmark (SPIVA, A)
0.0%
2020 年末顶部四分位美国大盘主动基金,4 年后仍在顶部四分位的比例——随机预期是 6.25%,一只都没有(A)Top-quartile US large-cap active funds still top-quartile 4 years on — random chance says 6.25%; the answer is none (A)
1.39%
中国研报「买入」评级的准确率(82.5 万份统计):「买入」占 37.87% 而最不准,「卖出」占不到 0.5% 而准确率 70.83%(B)Accuracy of «Buy» in 825k Chinese sell-side reports: Buy is 37.87% of ratings and least accurate; Sell is under 0.5% and 70.83% accurate (B)
510 亿
中国公募费率改革三阶段合计每年为投资者节省(元);综合费率降近 20%——费率单调下行不可逆(A)Yearly saving (¥) from China’s three-stage public-fund fee reform; blended fees down ~20% — the fee decline is monotonic and irreversible (A)
口径裁判:SPIVA(标普)与晨星「主动/被动晴雨表」方法论不同,中国 2024 年主动战胜率 13.4% 为资产加权口径,另有约 20% 的数量占比口径流传——引用必须带口径;流传最广的「GPT 财报预测胜过分析师(60.35% vs 52.71%)」论文 2025-02 已撤回,关联论文被期刊正式撤稿,凡以此为前提的推演一律重审;摩根士丹利 9.3 万亿为财富+投资管理合并口径,与分部口径差异巨大;散户端 AI 工具的量化效果本轮检索未找到任何证据——「散户平权」是推论,不是被测量的事实。Refereeing the numbers: SPIVA and Morningstar’s barometer differ in method — China’s 13.4% active win rate for 2024 is asset-weighted (a ~20% count-based figure also circulates); the famous «GPT beats analysts» paper was withdrawn in Feb 2025 and its companion formally retracted — every inference built on it must be re-examined; Morgan Stanley’s $9.3T is a combined basis; and no quantified evidence of retail AI-tool outcomes was found — «retail levelling» remains an inference, not a measurement.