Kyle Murphy

81 posts

Kyle Murphy banner
Kyle Murphy

Kyle Murphy

@kmurphy1

quant finance, econometrics, regression.

Beigetreten Mart 2021
165 Folgt111 Follower
Kyle Murphy
Kyle Murphy@kmurphy1·
At the bisection, methinks… Discovery without robust implementation is just research; implementation without ongoing discovery is just indexing. Reminds me of that saying “probability is the cleanest of mathematics and the messiest of life” (Taleb I think, but idk where he got it from or if its OC)
English
0
0
2
106
Markov's Chainsaw
Markov's Chainsaw@markovschainsaw·
hearing rumors that a pod is glowing up
English
1
0
1
31
Kyle Murphy
Kyle Murphy@kmurphy1·
@0xfdf @AgustinLebron3 Yes. Composites vs. primitives. My lean is that momentum can be one of the cleaner, higher-conviction primitives (12-1). Maybe longterm reversal? I think maybe in more HFT-like settings primitives could be preferred due to latency? But i just dont know.
English
1
0
2
299
fdf
fdf@0xfdf·
@AgustinLebron3 @kmurphy1 I think Kyle is describing that you usually want factors to be a function of more than one variable, so that each of the raw features captures a different characteristic of the factor. And then the question is - what factors does this not apply to? Momentum comes to mind.
English
2
0
9
655
Agustin Lebron
Agustin Lebron@AgustinLebron3·
Got a few hours on my hands. AMA.
English
68
0
112
27.3K
Kyle Murphy
Kyle Murphy@kmurphy1·
@__paleologo I feel they’re almost certainly running some “more advanced” (no idea what that might mean) versions of these models internally anyways?
English
1
0
1
1.2K
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
Honest question. Can Anthropic still run Fable/Mythos internally? I find a piece of alien technology in my backyard. That is the premise of a lot of scifi movies. If I were Amodei (or myself) first I’d grow a big of extra hair. Then I’d proceed to RULE MANKIND.
English
9
0
142
22K
Rev Culture
Rev Culture@RevvCulture·
Built to conquer any terrain, styled to look flawless doing it BMW R 1300 GS Adventure
Rev Culture tweet mediaRev Culture tweet mediaRev Culture tweet media
English
2
9
58
4.4K
Kyle Murphy
Kyle Murphy@kmurphy1·
It feels like so much of an allocator's problems rotate around the Marchenko-Pastur bound...but! The sample sizes allocators live with (modest list of assets and decade or two of monthly history) we see apparent patterns appearing by accident frequently. When we ask "how much genuine shared structure does my menu contain?" The math gives a precise cutoff for how strong some accidental pattern can be. What governs is less statistics and more a rulebook (hold benchmark unless a deviation is earned, stay inside a risk budget, cap positions, every weight explainable to a committee, taxes/fees and audit trail, etc.) The only place allocators get a sharp answer is where they least need one. The remainder is conventions and governance and what data can justify. Can this rulebook be written in the notation your books work in? Constraints, penalties, admissible sets. Can we design methods specifically for this seat? How much of this (tax lots, traceability, etc.) is irreducibly procedural?
English
0
0
0
25
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
@kmurphy1 We should maybe spend 1hr next time I travel to you or the other way around. Of course I’d like to do the right thing by allocators.
English
2
0
1
356
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
The Obizhaeva-Wang model is a standard model of market impact (linear impact, exponential decay). For fun, I checked if I could solve the optimal MVO problem with AR(1) and alphas. I knew it is solvable but it’s gnarly to write the state-space eqs, the Riccati eq etc. A few prompts, a little thinking, some checking, and it was done in one evening, closed-form solution, LaTeX write-up and everything. ChatGPT Pro. It could be easily extended to AR(p) and multi-exponential decay. Human in the middle, but that’s some real time saving right there. So you can think about the actually original models. This is nothing special, but it would have changed the way I wrote my second book (📕) and its topics, had it been available in 2024.
English
8
12
184
22.6K
Kyle Murphy
Kyle Murphy@kmurphy1·
I should be clear. "model-free" is a misnomer here. One must make an assumption on a factor model. You choose a lookback window for momentum. The weighting scheme is a choice. If this signal's turning point relocates depending on your choices then it's very much model-laden and exactly in ways that matter most.
Kyle Murphy@kmurphy1

Ehsani & Linnainmaa (2020) showed equity momentum as summation of factor autocorrelations. When you recover real-time factor returns fₜ and plot the momentum series as cumulative integral where the momentum-portfolio weights which are themselves a function of prior cumulative factor exposures... When that series stops going up or flattens and turns down, that's the point at which the net autocorrelation across all other factors as switched from positive to zero or negative. Whatever combination of value/quality/carry/vol that has been autocorrelated positively and was driving momentum stopped winning. That collective contribution to the integral peaked and the 'other factors' as a group have begun to mean-revert to their own recent path. One might conclude that the flattening of the momentum series is a clean, model-free real-time signal that the factor autocorrelation regime has exhausted itself.

English
0
0
2
131
Kyle Murphy
Kyle Murphy@kmurphy1·
R² = IC² R² is the squared correlation between forecast and outcome and IC is that correlation. A signal with IC=0.05 is monetizable and produces an R² of 0.0025. "3 features you know are predictive" and "terrible R²" are almost certainly both true. Maybe in ML an R²=0.02 is failure. For others it's a career (If they have breadth). In-sample ridge lowers R² relative to OLS. That penalty buys the variance reduction by sacrificing fit.
English
0
0
1
47
Markov's Chainsaw
Markov's Chainsaw@markovschainsaw·
@the_delta_dog are you looking at out of sample r2? how does it compare to unregularized regression? agree w lou / your comment that you should be looking at and comparing to ICs. Why not look at the IC of the regression output? r2 is arguably a garbage metric
English
2
0
1
491
DeltaDog
DeltaDog@the_delta_dog·
spent a week studying ridge regressions. up 3 hours before work today trying to learn and build an intuition in practice. used 3 features i know are predictive. trained many models, only to learn nothing and get terrible r2. unbelievable how hard this is to figure out on your own.. @quantymacro tips?
English
3
1
31
10K
Kyle Murphy
Kyle Murphy@kmurphy1·
Ehsani & Linnainmaa (2020) showed equity momentum as summation of factor autocorrelations. When you recover real-time factor returns fₜ and plot the momentum series as cumulative integral where the momentum-portfolio weights which are themselves a function of prior cumulative factor exposures... When that series stops going up or flattens and turns down, that's the point at which the net autocorrelation across all other factors as switched from positive to zero or negative. Whatever combination of value/quality/carry/vol that has been autocorrelated positively and was driving momentum stopped winning. That collective contribution to the integral peaked and the 'other factors' as a group have begun to mean-revert to their own recent path. One might conclude that the flattening of the momentum series is a clean, model-free real-time signal that the factor autocorrelation regime has exhausted itself.
sysls@systematicls

Remember that portfolios are linearly composable of other portfolios. This means that you can express a signal as a linear composition of other signals. This kind of composability is where most of the interpretability of modern quant finance comes from. The most fun signal to reason about is the momentum signal, because within the momentum signal contains everything that is “currently winning”. It is a dirty ensemble of all signals and factors that are currently positive. Individual stock momentum is largely subsumed by momentum in factors: factors autocorrelate, and sorting stocks on past returns implicitly loads you onto whichever factors recently won. Momentum is the only canonical signal that is a function of the other signals' realized returns. It is a meta-signal. It is also what makes it a good factor to neutralize against. You are implicitly removing the high variance factors that have recently done well.

English
0
0
12
2.7K
Kyle Murphy
Kyle Murphy@kmurphy1·
At risk of being the ultimate reply guy - which honestly idgaf. Such a passage, much literature. "The kids' horse sank beneath him, with a long pneumatic sigh"
Kyle Murphy tweet media
English
0
0
1
49
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
What is the hardest aspect of working in finance, esp. research? Not that things don’t work or research does not pan out 90% of the time. That’s research. Failure is a feature. Not the drawdowns. Being is a drawdown (sometimes painful or long) is the natural state of a strategy, and of life in general. The worst thing is the portfolio managers with whom you have a friendship of sorts who leave voluntarily or not. Sometimes, I even play a role in them leaving, which is both the right thing to do for investors *and* adds a healthy dose of sense of guilt. The weird thing is that I think of these people, even from a long time ago, *a lot*. Where are they? Could have we become real friends? Hey, I am not complaining, and maybe it’s the same everywhere (but slower). Samsara and anitya, or whatever. But I keep these verses of Raymond Carver in my memory: “Before long, before anyone realizes, I’ll be gone from here.” Good night, PMs.
English
13
6
200
22.9K
Kyle Murphy
Kyle Murphy@kmurphy1·
Acceleration in the age of cyborgs. We externalized thoughts first in drawings and pictures, then onto tablets with more formal writing systems. Is this a very large step further into externalizing thoughts and collaborating with not just one or others’ past selves - but some impossibly large artifice etched in silicon that is “all of our” past selves? What to do about collaboration with the machine spirits? Maybe Neil Postman’s “Amusing Ourselves to Death” discussion on the evolution from orality to writing to printing to television has something insightful to say on this. Each wave of the tech gives and takes. If writing weakened living memory, and printing gave us mass literacy with private conscience, what does the tit for tat that AI theorem proving at mass scale look like? Certainly like all waves passed it will change what we even mean by “thinking” and “knowing”, and to your point “communing”. If that’s always been humanity’s relationship to tech - is it a fair argument that there is nothing more human than that?
English
0
0
1
229
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
So true. Knowing that AI agents will prove or assist in proving theorems makes this classic image of two people, from different countries and times, thinking in silence about the same problem, in deep communion, incredibly poignant. And makes me afraid we are losing forever something ineffable that made us human.
Gabriel@gbrl_dick

i was reading through this drop last night and this photograph of erdos and terry tao had me tearing up. life can be so beautiful

English
8
14
245
27.1K
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
Short-ish rant: factor rotation, factor timing, factor horoscopes, skibidi toilet regression and many other quantitative malfeasances originate from quantitative "researchers" (both on the buy and sell side) whose incentive was to raise their status (and comp) by capturing the attention of investors. Because of the Bullshit Asymmetry Principle, dislodging these misconceptions is 10x harder than creating them. Ideally, incentive misalignments should be solved with contracts. But also, like in many professions, an ethical oath would help. First rule: "I won't bullshit the portfolio manager and the business owner alike even if doing so would benefit me". Feel free to like anonymously, I won't out you.
Gappy (Giuseppe Paleologo)@__paleologo

Once in a while I hear “Factor x is a bit weak in month y” and I die a little inside

English
18
14
378
68K
Gappy (Giuseppe Paleologo)
Gappy (Giuseppe Paleologo)@__paleologo·
These Balenciaga sneakers are for sale at a $1,200 price tag by the store next to my office, at the corner of 59 and Madison. They are probably worn by the target customers of the Ferrari Luce, for sale in the near future at the corner of 55 and Park.
Gappy (Giuseppe Paleologo) tweet media
Gappy (Giuseppe Paleologo)@__paleologo

Something must be done. This is a national tragedy. This is worse than the national team no longer qualifying for the World Cup. This is even worse than the death of Italian Cinema. Ferrari *is* Italy. Jony Ive can make slabs. He can't make cars. Give the job back to Pininfarina.

English
11
6
143
22.9K
Kyle Murphy
Kyle Murphy@kmurphy1·
So, the raw commission premium in Q1 is an aggregate. A good question is "how much of that premium can my optimizer harvest?" The first of some second-order diagnostics... (1/N) With some factor risk model you can decompose each stock's expected return into a factor component and residual component. Calculate your raw commission premium and substitute the factor decomposition. You can recover the factor-aligned commission with the benchmark-weighted mean factor exposure of Q1 minus the benchmark's factor exposure. See Ceria Saxena Stubbs (2012). See Gappy's red book. Ask yourself 'What's my residual commission premium?'
Markov's Chainsaw@markovschainsaw

@kmurphy1 - how to appropriately look at the signal neutralized to beta, vol, size, idio risk - equal vs cap weighted performance, and making sense of outliers (mean vs median). ie shld you trust a signal that only looks good on mean - trading off MCTE/diversification with signal strength

English
0
2
2
225
Kyle Murphy
Kyle Murphy@kmurphy1·
A book I inherited from Dr. Richard Ericson (Markov-perfect industry dynamics w/ Pakes). Serves to remind that optimization is tractable when the world’s compressed to the right inputs. Problem is whether the compression is valid!
Kyle Murphy tweet media
English
0
0
2
71

Entdecken

@markovschainsaw @0xfdf @AgustinLebron3 @__paleologo @RevvCulture @the_delta_dog @elonmusk @BarackObama