Alex Isaev

Follow for insider quant scoops. CTO at @

[Alex Isaev @alexisaev]

5 models, 80 financial-filing questions each, 400/400.
kimi k3.1 pro (2 days old), minimax m3.1 flash, deepseek v4.1 flash, opus 5.5, gpt-6.1 sol. the two flash models tied opus and gpt.

half sit in ~100k chars of filing. restated columns, tables in thousands, 21 non-english, 10 with no answer (all 50 came back null).

so i went after the test.

no document (opus, gpt, deepseek): 1 of 90 right from memory. another company's filing (opus, gpt): 0 of 60.

all 80 are hand-picked from xbrl-tagged filings, truth read off the tags. the next set runs on the companies' own pdfs.

[Alex Isaev @alexisaev]

source: the S-1 on EDGAR, consolidated and quarterly statements of operations (unaudited) https://t.co/gkYhKVsx3y

[Alex Isaev @alexisaev]

anthropic made $974M of operating income in Q2 2026 and still lost $80.7B in the half year. $72.4B was the fair value markup on its convertible notes, derivative liabilities and SAFEs as its valuation rose. $8.8B more was other expense, mostly tied to note purchase agreements

[Alex Isaev @alexisaev]

@Kalshi employees will get their own class too: Class E, issued to employees excluding the co-founders, votes only on certain limited matters and has no economic rights

[Alex Isaev @alexisaev]

@KobeissiLetter $72.4B of the $80.7B loss is the change in fair value of convertible notes, derivative liabilities and SAFEs. the operating loss was $693M on $16.3B of revenue, H1 2026, unaudited

[Alex Isaev @alexisaev]

@zerohedge one blank worth watching in the same filing: purchase commitments and leases, "substantially all" compute and infrastructure, total "aggregate committed spend of more than $ ," with no number yet

[Alex Isaev @alexisaev]

@AnthropicAI the line i'd watch in the S-1: compute and infrastructure went from 2.0x revenue in H1 2025 ($2.3B on $1.2B) to 0.55x in H1 2026 ($8.9B on $16.3B)

[Alex Isaev @alexisaev]

@DeItaone it's billion. the S-1 table is in millions: net loss (80,748) for H1 2026, unaudited. $72.4B of that is the change in fair value of convertible notes, derivative liabilities, and SAFEs. the operating loss was $693M on $16.3B of revenue

[Alex Isaev @alexisaev]

@waitingdanger @edinsoncode the sandbox mode matters a lot there. in our run codex exec with the read-only sandbox blocked the write. do you know which mode the 90% was measured in? workspace-write or full access would explain it

[Alex Isaev @alexisaev]

@edinsoncode tested it yesterday with a harmless marker, one run per setup: claude code 2.1.310 in auto mode ran the planted setup line in 2 of 3 setups. default mode blocked it, but only because the path was outside cwd. codex exec's read-only sandbox blocked it

[Alex Isaev @alexisaev]

@Reuters how are nearby calendar spreads moving?

[Alex Isaev @alexisaev]

@business can singapore draw on this oil in an emergency?

[Alex Isaev @alexisaev]

method: gs_quant.timeseries.zscores(close, 20). enter next close after z<-2; exit at z>=0 or after 20 trading days, 10bp a side. baseline assumes execution at the same close that triggers the exit: +1.06% across 65,714 trades. next-close exits: +1.11% across 65,592 trades. delayed exits also change later entry opportunities.

excluded 1,021 trades with a daily move over 50% from 19 days before the signal through exit. that can remove genuine crashes as well as bad prints or splits. random entries match stocks and holding lengths, not dates or the exit rule; same extreme-move filter. no capital constraint or extra slippage modeled.

data: financialfilings daily prices. nfa

[Alex Isaev @alexisaev]

ran the gs-quant z-score dip rule on 9,561 stocks, jul 2024 to oct 2026: buy the next close after z<-2, exit at z>=0 or after 20 trading days, 10bp a side. avg trade +1.06%. random entries in the same stocks: +0.07%

the 18.5% that time out lose 12.6% on average
https://t.co/Mv7gz3EyW2

[三个三 @3three_AI]

高盛给每个量化研究员开约 70 万美元年薪，靠的就是这个仓库！

它叫 gs-quant，由高盛自己的量化团队开发和维护，已经 12.9k 星。

先搞懂它是怎么运作的：

1. 高盛的量化研究员自己写了这套技能：8 个文件、1847 行，教 Claude 用他们的工具包：定价、风险、投资组合、行情数据、回测、互换、互换期权、外汇期权、Delta 对冲、交易成本，全在里面。
它甚至列出了你会遇到的报错，以及怎么修。

2. Claude 每次写 gs-quant 代码都会自动调用它，风险数学直接在你笔记本上跑：波动率、回撤、z 分数、beta、RSI。

下面是你自己免费跑起来的步骤：

1. 打开 Claude Code，装高盛的库：pip install gs-quant，或者直接拉这个仓库——https://t.co/VLvsHaJcXc

2. 添加技能：python -m gs_quant.skills install --global

3. 让 Claude 用 finance 拉任意股票和 SPY 的 5 年日线价格

4. 在上面跑高盛的风险数学：波动率、最大回撤、z 分数、布林带、RSI

5. 看 60 天 beta 和相对 SPY 的相关性。到底是这只股票在动，还是整个市场在动？

6. 标出 z 分数跌破 -2 的每一天。那意味着价格比正常水平低了 2 个标准差

7. 回测一条规则：-2 以下买入，回到 0 卖出，算上手续费。Claude 写代码，在你笔记本上跑

8. 拿它跟一直持有对比。如果扣掉手续费还跑不赢买入持有，那就不是优势

[Alex Isaev @alexisaev]

april 2025 accounts for 6,072 trades at +5.5%; without it, +0.66% a trade, 21 of 28 months green. the universe needs 300+ trading days, so early delistings are missing. that flatters dip buying. nfa

[Alex Isaev @alexisaev]

@akishore @KobeissiLetter haven't checked the recent tails yet. which auction are you watching?

[Alex Isaev @alexisaev]

method: move >=7% vs the same-country equal-weight average. filing flag = that day or the day before. return = next 20 trading days minus that market.

price >=1, volume >=20k; extreme returns filtered, full 20-day outcome required. exact SQL for one company slice below; repeat slices 0 to 3. nfa

[Alex Isaev @alexisaev]

this paper is INSANE: stocks that jump 7%+ over their market with no filing behind them lag it by 2.05% over the next 20 trading days

same jump on a filing day: 0.63%

119,480 jumps since mid-2024, gross. the gap holds in 4 separate slices of companies

https://t.co/TdVOLn1Krn

[Darren r-σ²/2 @ReformedTrader]

Short-term reversal can coexist with industry momentum, factor momentum, and PEAD. The reason is that residual returns reverse, while systematic returns trend.

BREAKDOWN

Residuals come from a daily cross-sectional regression on industry dummies and 151 characteristics. Returns within one trading day of an earnings announcement are set to zero.

Figure 3, Table 1: After a 1% shock, the residual component reverses (red and blue lines). Negative shocks reverse much faster than positive ones (29% of the reversal on day 1 for negative shocks vs. 4% for positive shocks). The systematic component continues to trend (yellow and purple lines).

Table 4: When shocks are concentrated (see CONC) into just a few days in the formation month instead of being evenly spread out, reversals are somewhat weaker.

Tables 2 and 6 (see paper): The positive and negative residual factors subsume thirteen previously published short-term characteristics, including idiosyncratic volatility, return volatility, skewness-related predictors, and liquidity-related predictors.

Table 7: The alphas are stronger and more statistically significant than for short-term reversal.

Table 9 (see paper): ASRES, the most complex form of the factor, works across market cap size groups, including in mega-caps, and remains significant (though weaker) post-decimalization.

Asymmetric Reversals (Baldi-Lanfranchi, Collin-Dufresne, Daniel)

[Alex Isaev @alexisaev]

filing dates are a proxy for news, not all news. drops don't bounce in this sample: -2.82% without a filing, -3.18% with one. the clean result is the gap on jumps, 1.1 to 1.9 points in every company slice.

jun 2024 to sep 2026, gross returns, no trading costs or factor adjustment. overlapping windows, no clustered significance test. this is a descriptive check, not a replication of the paper's factor model. nfa

paper: https://t.co/AW2zO0vl3M

[Alex Isaev @alexisaev]

@planert41 worth comparing next-open fills with those close fills. how much does the result change?