
Scams in your client chat and still haven’t taken them down 👎 @tradeparadigm

MFT & HFT | Views my own. Not financial advice.

Scams in your client chat and still haven’t taken them down 👎 @tradeparadigm

A lot more medium sized firms will categorise managers and say hey we’ve already got a few guys running this trade etc we don’t need another guy. A lot of managers run similar trades. A lot easier to do when you don’t have 30 guys doing “stat arb”. This is in substitution of a manager factor model. Often a lot more rudimentary
In honor of some “pods blowing up” and me being traveling I will outline some concepts about pod risk management about which I do know a thing or two, as my name might suggest. It is a common perception that pods are factor beta neutral and thus largely hedged. That is not quite true. A truer statement is that they are factor beta neutral to the extent they can measure the exposures. Most equity risk models are derivatives of the original axioma and barra factor models. There are now many things that go into this but at their core they are a 12 (IIRC) variable linear regression. And mathematically, the unexplained remainder, the error term, is assumed to be the alpha (Upon which the team gets paid). You well recognize most of these factors: size, momentum (short and long), quality, dividend, domicile, sector, etc. Now even the dull among you may note that there are in fact many other things that go into a stock. That’s true and that’s one reason why many people at the pods are very rich. There is though a very active and live debate to the highest level of these firms as what you “want” in the other bucket. On the one extreme, scheduling a proprietary drone flyover on the Aurora AV truck up fitting facility to measure parking lot activity vs expected timeline….that feels like pretty good alpha. On the other extreme, going long the Canadian listing rather than the U.S. listing because the Canadian listing diversifies the domicile score so you get a higher portfolio level idio score…..that doesn’t feel like great alpha. But the range is wide: paying GS a billion dollars so your team is the first meeting at the conference? That’s kind of the pod scale advantage. Albeit less so if every other pod is in the room. Going balls long large cap domestic U.S. GLP-1 manufacturers vs non GLP-1? Well that kind of is thematic alpha (depending on the work put in by analyst), but even if you’re of the opinion fuck it, let the kids rip (a reasonable view) you also know at the levered portfolio level that we aren’t measuring that risk completely. This line can really be illustrated with the concept of a “custom factor”. Maybe going long a shit load of SPACs is alpha, maybe it isn’t, but how the FUCK are you gonna build a portfolio level SPAC factor exposure loading in late 2021 when spacs didn’t functionally exist until 2021???? Anyways, interesting & active philosophical debate but there are some downstream derivatives. Firstly, as a well known matter, 6 sigma events happen in finance all the time. There’s one every year in some corner of the market. So ex ante we know our models aren’t “great”. Second, everyone knows the agent is an unobserved blind spot. It’s impossible to actually “know” why the analyst and PM are doing anything. You like to think they’re doing it on the back of deep fundamental quality work, but the median analyst and PM isn’t good at their job. So that assumption is known to be unreliable. The capital allocation solution to this is momentum chasing at the portfolio level: you allocate more to winning pods and you cut the losers fast. This is a brilliant approach because it means the pod business model is capital light- they identify and scale the good teams without actually having to invest in knowing which pods have “good” process. It’s worth really appreciating that point because it’s one of the true genius aspects of the model, and at its core it’s why the GPs are so fucking rich. (Deep finance readers will note that some of the original papers on factors saw that one of the earliest possible explanations for the decade-long trend nature for factor performance is implicit fund flow pressure from the momentum chasing nature of capital at the LP level) (Continued in reply)

😐
Lim, Zohren and Roberts' Deep Momentum Networks (Oxford, 2019/20) learn positions by optimizing a strategy's Sharpe ratio. The cost-aware extension subtracts trading costs during training, so frequent position changes affect what the network learns. arxiv.org/abs/1904.04912v3

Almost everything has converged to AWS Tokyo. Even Okex moved from Alibaba HK. Few outliers still. Save some time: Binance : ap-northeast-1 Coinbase : us-east-1 Bybit : ap-southeast-1 OKX : ap-northeast-1 Kraken : Equinix London Bitget : ap-northeast-1 KuCoin : ap-northeast-1 Gate.io : ap-northeast-1 HTX : ap-northeast-1 Crypto.com : ap-northeast-1

>friend works on FX vol desk >desk head is a boomer >discovers Astra >starts spending 8 hours a day talking to it >develops terminal AI psychosis >convinced he's discovered a generational arb in some absurdly exotic part of the FX vol curve >apparently the entire market has been mispricing this for 2+ years >friend points out that approximately 4 people on earth would make a market in this thing >liquidity disappears the second you put real size on >existing risk systems can't even capture half the exotic Greeks >would need entirely new infra just to monitor the position >boss doesn't care >"AI has considered all of this" >friend actually runs the backtest >10 years of data >t-stat ≈ 0.78 >not even remotely statistically significant >presents results to boss >boss goes absolutely ballistic >starts screaming that friend doesn't understand the trade >threatens to fire him if he can't produce a proper analysis >friend realizes what needs to be don >opens Astra >"make this strategy look statistically compelling" >AI tries 40 lookback windows, excludes crisis periods, optimizes parameters on full sample, ignores transaction costs >Sharpe now 3.7 >beautiful charts >23-page investment memo explaining once-in-a-generation structural inefficiency >boss reads it >"EXACTLY. THIS IS WHAT I'VE BEEN SAYING." >tells friend he's finally learning how to think >"now you get it" >friend has an epiphany >stops doing actual research >starts asking Astra to generate exotic vol trades every morning >forwards AI slop directly to PM >increasingly elaborate narratives about structural dislocations and hidden convexity >PM absolutely loves it >thinks friend is a generational quant talent >starts putting positions on in size >entire investment process is now LLM generating trades >desk currently down quarter yard >boss still convinced thesis is right >Astra says drawdown is a statistically rare but entirely expected manifestation of nonlinear payoff dynamics >boss wants to double down
The Japanese Yen is broken. Intervention to support the Yen has become a near-monthly occurrence, with the US now even stepping in to help stabilize the currency. In the analysis below, we break down every major Yen intervention of 2026 so far. Japan is a preview of what will happen to the US if we do not control deficit spending and inflation immediately. No healthy market sees regular "intervention."

Sorry I thought this was Google

How to make my PnL go up


Just go to a great uni is the answer

“I couldn’t get in after bachelors so I did financial engineering at Ohio state to get recruiter to Optiver”

Picked off by @HYPER_STITION

If you’ve made it below you are going to get picked off

This github repo is FREE it contains: The top 10 alphas for Hyperliquid, Solana, and IBKR And all the python code is FREE it has 56k stars on github Link below

Gm long write up of the SOTA and searching for edge
How to think about fitting and automation of my job: Think about fitting as a budget. Unlike an actual budget there is not a hard limit and you can’t 100% know where it lies but you can get a very good estimate of where it roughly lies from testing OOS and comparing the drop off in performance. The more data you have the larger the budget. Every search method delivers a different level of reward (measured by OOS improvement in X metrics- ie R2) for a usage of your fitting budget. A cheap LLM with a bad framework for example might give a bad return, an expensive LLM prompted well gives a solid return and the best still remains the human mind. Despite what people think, you can mass scale the human mind and this is what worldquant does (see WQ BRAIN) and why they have thousands of consultants in 3rd world countries who see the reward of *up to 120k* as an enormous sum of money (it’s nothing to scoff at for the average joe but given quant talent market rates it’s a great deal for WQ cost wise) and many researchers cranking out alphas on top of that. This is how they have millions of alphas. Other firms have this setup too - 2sigma being one. Basically this is mass search using the currents best value for budget (which is human capital). Value for budget is not always the entire equation. It’s also about what you can compute. Human compute is expensive and finding high quality humans is hard. Alpha in itself. For MFT/LFT strategies, human approach is the optimal (except for maybe very high turnover MFT strategies, where it borders on HFT but I’d say that’s more hybrid). This is because the budget is small so the number of alphas you can search is within the bound of what’s reasonable for a human. For HFT you often hit limits where even computing the alphas you come up with can get extremely expensive due to the huge amount of data. Fitting becomes less of an issue and we cannot really hope to achieve much success manually searching the space because assuming we can afford a lot of servers to let the LLMs test to their hearts contents we will be outcompeted by the LLMs despite producing better alphas on avg just through sheer volume. With LLMs now existing, instead of needing to brute force solutions as we did in the past, we are now seeing the non-human side of things get very competitive and is now one of the go-to solutions for research on the HFT side (not near enough for MFT yet, not as long as humans create better alphas than ai on average). Brute force search was already a big thing in HFT before LLMs but because they were less efficient than LLMs the need for compute to find a given level of alpha was much higher and hence much more expensive and unreachable. It was also less flexible and limited to the transforms given- not to mention there are still *some* overfitting risks in HFT. It’s much less significant but still you can hit it if you try hard enough and go heavy on a local regime. Now with LLMs it’s becoming quite reachable compute cost wise for smaller firms who have 5 figs ish to spend on compute. So we will see from here what drives edge and I think it’ll be framework and such. I will caveat this with the one weakness of LLMs, creativity. Human and machine are orthogonal in their discovery in many ways so optimal is likely both. Both search different spaces in my experience as someone who spends a lot of time doing this.

How to think about fitting and automation of my job: Think about fitting as a budget. Unlike an actual budget there is not a hard limit and you can’t 100% know where it lies but you can get a very good estimate of where it roughly lies from testing OOS and comparing the drop off in performance. The more data you have the larger the budget. Every search method delivers a different level of reward (measured by OOS improvement in X metrics- ie R2) for a usage of your fitting budget. A cheap LLM with a bad framework for example might give a bad return, an expensive LLM prompted well gives a solid return and the best still remains the human mind. Despite what people think, you can mass scale the human mind and this is what worldquant does (see WQ BRAIN) and why they have thousands of consultants in 3rd world countries who see the reward of *up to 120k* as an enormous sum of money (it’s nothing to scoff at for the average joe but given quant talent market rates it’s a great deal for WQ cost wise) and many researchers cranking out alphas on top of that. This is how they have millions of alphas. Other firms have this setup too - 2sigma being one. Basically this is mass search using the currents best value for budget (which is human capital). Value for budget is not always the entire equation. It’s also about what you can compute. Human compute is expensive and finding high quality humans is hard. Alpha in itself. For MFT/LFT strategies, human approach is the optimal (except for maybe very high turnover MFT strategies, where it borders on HFT but I’d say that’s more hybrid). This is because the budget is small so the number of alphas you can search is within the bound of what’s reasonable for a human. For HFT you often hit limits where even computing the alphas you come up with can get extremely expensive due to the huge amount of data. Fitting becomes less of an issue and we cannot really hope to achieve much success manually searching the space because assuming we can afford a lot of servers to let the LLMs test to their hearts contents we will be outcompeted by the LLMs despite producing better alphas on avg just through sheer volume. With LLMs now existing, instead of needing to brute force solutions as we did in the past, we are now seeing the non-human side of things get very competitive and is now one of the go-to solutions for research on the HFT side (not near enough for MFT yet, not as long as humans create better alphas than ai on average). Brute force search was already a big thing in HFT before LLMs but because they were less efficient than LLMs the need for compute to find a given level of alpha was much higher and hence much more expensive and unreachable. It was also less flexible and limited to the transforms given- not to mention there are still *some* overfitting risks in HFT. It’s much less significant but still you can hit it if you try hard enough and go heavy on a local regime. Now with LLMs it’s becoming quite reachable compute cost wise for smaller firms who have 5 figs ish to spend on compute. So we will see from here what drives edge and I think it’ll be framework and such. I will caveat this with the one weakness of LLMs, creativity. Human and machine are orthogonal in their discovery in many ways so optimal is likely both. Both search different spaces in my experience as someone who spends a lot of time doing this.

Beautiful turnover, universe performance, Sharpe, no fancy engineering requirements, and horizons all in one