JJCoast

[Jérémie Boyer @jeremie0117]

Duquesne Family Office
(Stanley Druckenmiller Fund)

[Livsun @L1vsun]

citadel and two sigma pay millions for positioning data. same data is free on a government website, updated every friday since 1986

it's called the commitment of traders report, published by the CFTC

every week: exact net positions of commercial hedgers, large speculators, and small traders across every futures market on earth

commercial hedgers are the ones who actually move price - producers, banks, institutions trading to hedge real exposure

they've been net-short before every major equity crash in the last 30 years

when commercials flip heavily net-short while large speculators go max long, that setup has resolved bearish 74% of the time on a 6-week window

that's not a prediction. it's reading a positioning table published for free every single friday

quant desks run this as a regime filter - if the report shows big money hedged against your direction, you're buying into an invisible wall

data goes back to 1986, sits on cftc. gov, and takes maybe 2 hours in python to clean

you've been reading the chart, which is the outcome of this positioning

save this - you should be reading the cause

[Didier Lopes @didier_lopes]

7 years ago I started my "financial career" by forecasting the S&P 500 with LSTM RNNs with tensorflow.

LSTMs were still fairly new back then.

I spent a lot of time thinking about each of the hyperparameters (activation function, number of layers and neurons, training epochs, dropout rate, optimizer).

Recently I wanted to see how far the field has come. So I tried @nixtlainc , which my friends Max and Cristian have been building for the past 5 years. And also backed by @Microsoft !!!

One of their most advanced models is TimeGPT-2, a transformer-based foundation model trained on over 100 billion data points. You point it at a time series, pass a few parameters, and you get a forecast. The API even does embedded fine-tuning.

Crazy how far I can get with a couple lines of code, what took me weeks in the past ....

So, obviously I applied TimeGPT-2 to a few macro data series and the experience was seamless. It feels like at some point we might have Forecasting as a Service lmao

Anyway, I made the entire experiment open source, and it runs as an @openbb_finance App that you can add directly to the OpenBB Workspace. The recipe is also available so it's easier to extend to different timeseries.

Let me know what you think!

[Brent Sullivan @TaxAlphaInsider]

If an adviser wants access to long/short strategies for an ultra-high-net-worth client, but Fidelity and Schwab have restricted their access, they have three options:

1. Join up with a bigger practice
2. Use another custodian like Interactive Brokers, Goldman Sachs, and Pershing
3. Trade directly with prime brokers like Morgan Stanley, JP Morgan, and Goldman Sachs.

To trade with prime brokers, an investor typically establishes a legal entity to house the assets, usually a limited liability company (LLC) or a limited partnership (LP).

The wrapper makes a big difference in how tax benefits ultimately flow.

My latest blog post is about the pros/cons of LLCs and LPs in the context of long/short strategies.

Link in profile.

[fdf @0xfdf]

Another hangup I'm seeing: why did Jane Street have long momentum exposure? Not because they're yoloing a factor. Because many stat arb strategies organically accrue momentum exposure and residualizing it out would reduce risk-adjusted returns.

A stylized example with no IP since it's in the literature: your signal is xsection(analyst earnings revision). Long equities with upside revision acceleration, short equities with downside revision acceleration. This is _not_ a bet on price continuation/trend.

However! Analysts empirically cluster around news that has already moved the price. The legitimate fundamental information seeding the cross-section has a latent loading on momentum, via human behavior.

If you try to orthogonalize this signal to momentum, you linearly downweight the equities in the cross-section which report the highest analyst conviction, because that conviction is mechanically reflexive on momentum, even though it is a qualitatively distinct source of information!

Consequently, your residualization will flatten the distribution: post-resid top quintile of the xsection has shifted names with weaker signal from the middle three quintiles, and likewise for the bottom quintile.

Realistic numbers - you squeeze a 1.0 Sharpe out of analyst revisions, 8% annualized on 8% vol. You residualize out momentum, your vol does drop because you took out the momentum risk. Maybe it goes to 6%. But oops - because of the distribution smearing effect, your returns have fallen faster, to around 5%, and now you've got like a 0.85 Sharpe.

Jane Street are not idiots. They understand this intimately well, and far better than I'm even describing here. They're still up $40B net revenue YTD after losing $15B because they rationally accept the higher risk adjusted returns that keep in the factor prone to sharp drawdowns.

[Gappy (Giuseppe Paleologo) @__paleologo]

I finished reading “The strangest Man” a couple of days ago and am still thinking about it (and Dirac’s life and beliefs); which, to me, is the mark of a good book.

A movie about Dirac would be far more interesting (if not more commercially successful) than “Oppenheimer”. And why don’t we have a decent biopic of Newton, or Hilbert, or Gauss? These were genuinely interesting people! “Wittgenstein” by Jarman was just a collection of known sketches. “I Ragazzi di via Panisperna” (about Fermi) was excellent. “The man who knew infinity” (about Ramanujan) was also very good (Patel and Irons could make a list of groceries interesting). “Death of a Neapolitan Mathematician” (Caccioppoli): also great, if you can find it.

We need far more biopics of scientists and philosophers. The mix of mundanity, humanity, and their unique obsessions. Make scientists cool again.

[Livsun @L1vsun]

a physicist at bell labs accidentally solved trading in 1956

he wasn't trying to. he was working on information theory for telephone signals

the paper sat in an academic journal for two decades. renaissance technologies found it, wired it into their entire operation, and ran 66% annual returns for 35 years without a single losing year

it's called kelly criterion and it answers the one question no indicator on earth can touch: how much do you risk on any given trade

retail obsesses over entries. quant desks figured out decades ago that entries are maybe 20% of the edge

edge:
f = (p * b - q) / b

p = win rate
b = reward-to-risk ratio
q = 1 minus p

55% win rate on 2:1 trades → kelly outputs 32.5% of your risk pool per trade, not your full account

the difference between those two things compounded over 300 trades is not subtle

run this on any system you've backtested and something uncomfortable shows up - most retail "edges" fail the math at 50-60 trades. the signal was fine. the sizing destroyed it

quants don't guess size. the formula runs. the formula decides. no override, no feel, no vibes

the paper is free on google scholar. the implementation is 8 lines of python

Bookmark it

they had you studying rsi divergence for 10 years while the actual variable sat in a 1956 physics journal

[Peter - Cracking Markets @SystematicPeter]

Many of you have sent me messages saying my implied-volatility mean-reversion framework inspired you to build, and now trade your own variations.

That is exactly why I published it.

MRIV has continued to perform well in the post-publication backtest:

• Sharpe: 2.24
• Outperformed SPY over the same period
• Average exposure: only 6.51%

Many stock mean-reversion systems time entries using realized volatility.

MRIV uses implied volatility instead.

The general logic:

• Compare today’s decline with the range implied by volatility
• If the decline is unusually large, look for another drop the following day
• Use implied volatility to define the expected range and profit target

This makes the rules adapt to how the options market is pricing each stock’s current risk.

When implied volatility is already high, even a deep decline may not be exceptional. The expected range expands, the entry hurdle rises and there may be no trade.

Simple principle. Adaptive rules.

I download the required implied-volatility data free through IBKR.

Full deep dive + daily updated statistics:

https://t.co/NC1UpC3Q5M

[venus @RitOnchain]

A new paper from Stanford, Princeton and Ant Group introduces AQuA, an autonomous quant research agent.

To stop the AI from accidentally looking into the future, they locked it in a "sealed sandbox."

The resulting threshold long/short strategy generated a +2.50 Sharpe ratio out-of-sample.

Structural constraints beat prompt engineering every single time.

Bookmark and read the full paper.

[爱吃折耳根的Ace @_zheergen]

全球最大对冲基金之一的 Man Group 把自己的数据库引擎开源了

兄弟们！这个 DataFrame 数据库是专门给量化设计的

Man Group 开源的 ArcticDB，是一个高性能、无服务器的 DataFrame 数据库，专为 Python 数据科学和量化场景打造，目前有 2.5K Star。

核心特点如下：

1️⃣ 直接读写 Pandas DataFrame，用起来很自然
2️⃣ 专为时间序列优化，能处理数十亿行数据
3️⃣ 无服务器架构，数据可存 S3、LMDB 或 Azure
4️⃣ 支持版本回溯（Time Travel）和快照
5️⃣ Schema 灵活，可随时追加和修改
6️⃣ C++ 引擎加速，查询和处理速度很快

对需要管理大规模行情、因子或回测数据的量化研究者来说，这个工具解决了“数据存取慢、管理麻烦”的痛点。

以前存行情数据：CSV 存爆硬盘 → 换数据库 → 查询慢成狗 → 放弃。现在 ArcticDB 几十亿行行情数据本地就能跑。量化人的"数据仓库"，终于有正解了。

[JJCoast @jjcoastquant]

@benln Where is it? Looks like NOHO to me

[Christina Qi @christinaqi]

It actually blows my mind seeing this new wave of hedge fund startups raising GP funding instead of, you know, LP funding…

[App Hunter  @Sn0wbrave]

https://t.co/stM8v8rvp7

[JJCoast @jjcoastquant]

@itszyven Yes

[SK @StepOneAi]

私信有人问我关于优化 Funding Rate 交易信号

在构建基于 Funding Rate的交易因子时，单一指标往往容易产生噪音。要提高因子的有效性，需要进行结构化处理与多维交叉验证。

不同币种的 Funding 基线存在结构性差异。零售资金偏好的币种常年维持高费率，单纯进行横向绝对值比较，衡量出的往往是“散户化程度”而非“真实拥挤度”。

建议可以先对单一币种的滚动历史数据进行 Z-score 标准化，剥离基线差异后，再进行横截面 Rank 排序。

还可以引入 OI (持仓量) 维度的交叉验证
相同的高费率状态，市场结构可能完全不同，前瞻收益预期也截然相反。必须结合 OI 的变化来判断：

funding 高 + OI 上升，是有人在新开多单,拥挤在积累。
funding 高 + OI 下降，是老仓位在平掉，拥挤在出清。

还有加入价格维度可以进一步提高信号胜率。若价格处于横盘或阴跌状态，但 Funding 仍维持高位，通常意味着多单正在支付高昂成本被动扛单。这种“量、价、费”结合的组合信号，指向性显著强于单一维度。

同时Binance、Bybit、OKX 和 Hyperliquid 等平台的费率存在微观差异。更严谨的做法是使用 OI 加权合成全局 Funding 指标。此外，各大所费率的离散度（分布差异）本身也是一个极佳的情绪信号。
这类拥挤度因子的衰减速度较快，对于日频以上调仓的 CTA 策略，可能难以有效捕获此类信号，更适合中短周期的波段策略。

[JJCoast @jjcoastquant]

@james406 Not true. Chromebook is just a terminal to google servers with TB memory

[JJCoast @jjcoastquant]

@arhamamiin Luxury of being able to work on your interest

[Sonya Huang 🐥 @sonyatweetybird]

the smartest young people in sf were working on agi/alignment 5-10y ago.

they are working on bci today.

naomi is a total star and is one of an explosion of young talent into the bci space recently. i’m very excited about what that means for the world in a decade. 🧠

[JJCoast @jjcoastquant]

@je_capital @negligible_cap Is there Asian equivalent of Bloomberg?

[Voltex @VoltexGar]

The smartest player always loses to the one who owns the rules. a Stanford professor proved it. Wall Street built a $20 billion machine on it.

it is called algorithmic game theory, free on YouTube for years. the professor is Tim Roughgarden, and his whole point is one move. stop trying to out-play the game. sit in the seat it was built to reward, and let everyone else supply your edge.

the post above is that seat, in cash. Jane Street does not predict the market. it takes a toll on a third of every retail order in America and books $20.5 billion whether you win or lose.

the mechanism does not care how smart you are. it margin-called Bill Hwang and his $36 billion into 18 years, in two days.

that same lesson runs Google's ad auction, the traffic on your commute, and every marketplace you have ever paid a fee to. the money is never in playing well. it is in owning the rules of play.

no prediction. no genius. one seat, and the rules.

the smartest player is in prison. the house that wrote the rules is up $20 billion. the difference is one free Stanford lecture.