fdf
@0xfdf
Followers
10,887
Following
83
Media
71
Statuses
664
Wir müssen wissen, wir werden wissen.
Chicago, IL
Joined September 2023
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Wizkid
• 227500 Tweets
Davido
• 196578 Tweets
WONWOO
• 122420 Tweets
からくりサーカス
• 103453 Tweets
ケンタッキー
• 100048 Tweets
FAYEYOKO BTS 1ST FANMEET
• 98042 Tweets
BTS NUNEW 1ST CONCERT EP2
• 90630 Tweets
ニンテンドーミュージアム
• 83493 Tweets
#青木宣親
• 67013 Tweets
Supreme God Kabir
• 46362 Tweets
青木選手
• 41910 Tweets
ヤクルト
• 37400 Tweets
青木さん
• 27710 Tweets
Alcaraz
• 26135 Tweets
Cassper
• 17927 Tweets
Shana Tova
• 17565 Tweets
ラムジェット
• 17441 Tweets
サンライズジパング
• 16698 Tweets
TORE
• 13495 Tweets
ダービー
• 13189 Tweets
スワローズ
• 12584 Tweets
שנה טובה
• 12477 Tweets
I'd like to endorse
@__paleologo
's forthcoming book, The Elements of Quantitative Investment. I've now read most (but not all) of the draft chapters and I think it's probably now the best single, modern introduction to the field. There are also things for practitioners 🧵.
5
26
417
Very good thread. To extend this: signals that have positive correlation with future returns are still useful even if they don't surmount transaction costs on their own, because they lower volatility and costs when combined. I will show this via simulation.
Correlation between your signal and future returns is an important metric in quant trading. But what is a “good” correlation? Here’s a simple way to think about it.
13
99
964
6
24
337
If you're wondering how to derive the (annualized) Sharpe from win rate and n bets:
S = (2 x win_rate - 1) x sqrt(252 x n_bets)
Note that n_bets is the number of _independent_ bets, and this is on idio returns. If you're market/factor neutral, this calculation works.
I mean if you happen to have a ~6 Sharpe strategy like the simulation here you shouldn't rly have problem focusing on the big picture
7
3
199
14
27
319
For calibration:
1. I expect a junior quant candidate to nail all of these. These are all undergraduate linear algebra.
2. These are not brainteasers nor mental arithmetic.
3. None of these should be answered by talking about code or a specific library/language.
@macrocephalopod
"For each of the following questions, describe your answer as rigorously as you can.
Question 1: I tell you that I am regressing Y on X. Describe what I probably mean as rigorously and completely as you can, without sacrificing generality.
5
14
203
9
18
310
This paper is unserious for any real practitioner, and I will show why with citations.🧵Along the way, hopefully this will be an example of how to quickly scan a new "strategy" or "alpha" paper for credibility.
For posterity, the paper is here:
New Trading Ideas! How to construct time-series momentum portfolios with better risk-adjusted performance?
The paper "Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning" proposes a novel approach using MTL.
Keep Reading! 👇
1
12
72
10
22
224
A🧵on pricing data. I'm currently working on a series of posts studying different approaches to portfolio optimization. Since this is extracurricular, I won't use my firm's equipment or data (so: no bloomberg or factset). But "high quality" alternatives display trivial errors.
6
12
221
Outsiders believe the most successful people in this industry are the best at math. My experience is that the most successful are actually just "good enough" at math, can compartmentalize what they don't know effectively, and are extremely creative.
4
5
208
@macrocephalopod
"For each of the following questions, describe your answer as rigorously as you can.
Question 1: I tell you that I am regressing Y on X. Describe what I probably mean as rigorously and completely as you can, without sacrificing generality.
5
14
203
Maiden Century is a good platform, but there are more clever ways to construct signals from typical sources of alternative data. Here is an example. Let's say you have some agnostic alt dataset that looks like this plot. How would you build an expected return signal from this?
There is an alt data tool gaining traction called Maiden Century which I find to be quite interesting. Not because it does anything that novel but because of the implications of what it will mean when there is widespread use of it at the non-MM HFs
3
0
46
3
21
202
For students and aspiring quants, this is a good thread:
- first person recommends a paper
- another clarifies the intended use
- a third replies with a conjecture
- that conjecture is disproven
- the disproof is critiqued with a practical aside
All involved are professionals.
Modern ridge regression had been adapted to factor pricing models and detecting signals in a number of papers. For very technical reasons, I don’t find this technology transfer satisfying. There are regime gaps and strong assumptions in the original papers, and the finance papers
7
15
233
2
13
188
What is the *best* argument you can make that a hedge fund should *not* be market neutral? Market neutral is a low bar, and doesn't include sector + style neutral more generally.
49
6
164
@quantymacro
The asker's premise is slightly incorrect: we don't necessarily desire orthogonal matrices, but rather matrices with orthogonal column vectors (the former is a stronger notion than the latter).
Orthogonality itself is desired because it means the features the column vectors
2
7
143
Mosek's portfolio optimization cookbook is kind of great. I feel like if you get a good grasp of the underlying math from Vandenberge-Boyd, this is a concise, practical and comprehensive survey of portfolio optimization in practice.
3
12
138
This is incorrect. It would be correct if it were talking about ordinary returns, where the vol does scale with the square root of time (hence annualizing Sharpe by multiplying by sqrt(252)).
But compound returns are not a stationary process; the vol depends on the time period.
7
2
130
This is the start of a good thread on the Kelly criterion. Areas of discussion: is it possible to rigorously defend fractional Kelly? And what are minimum conditions for log optimality?
I thought Kelly optimality required identical likelihoods, but I was mistaken!
@__paleologo
Kelly is an idea perfect for stationary games, like blackjack.
Plus, Kelly is suboptimal when 3rd and 4th moments are non normal (chart below).
And good luck estimating moments so accurately out of sample!
But as a guide… sure, I can think of worse.
2
6
50
3
6
126
This paper is trivially invalid.
Don't need a long thread, just start with checklist item
#1
: impact. On page 24 they disqualify their results by estimating a bespoke "slippage" model on $100k SPY orders. Of course they found approximately no impact.
Is this real?
“Beat the Market: An Effective Intraday Momentum Strategy for S&P500 ETF (SPY)”
16
6
120
9
3
119
A sobering question: suppose you estimated an annualized Sharpe of 2.0, from annualized return 0.1 and annualized volatility 0.05.
How many days of trading do you need to obtain a margin of error of only 0.1 on your Sharpe, at the 95% confidence interval?
If you're wondering how to derive the (annualized) Sharpe from win rate and n bets:
S = (2 x win_rate - 1) x sqrt(252 x n_bets)
Note that n_bets is the number of _independent_ bets, and this is on idio returns. If you're market/factor neutral, this calculation works.
14
27
319
9
2
100
For a reference, I like _The Sharpe Ratio_ by Steven E. Pav. Approximations of the Sharpe standard error differ, but this falls out of S +/- 1.96 * sqrt(SE(S)). Usually SE(S) is about (1 + tau S^2) / t, where t is the number of time periods (NOT trades), and tau is ~0.5-1.0.
There is no way to estimate SR of 18 to the 4th significant digits. It would take 1M years…
4
4
116
1
4
82
The
@stripepress
reprint of Hamming's The Art of Doing Science and Research is really quite nice. "You and Your Research" is the crowd favorite, but I find myself coming back for the rest of it too. Hamming surveys vast amounts of land without getting lost in the forests or mud.
6
4
84
@macrocephalopod
The duality of traders:
- Trading is the most competitive game in the world, most fail, it lies at the bleeding heart of our capitalist society and we are glatiators raging against entropy
- Hey man I like money, I can code and I get how regression works
2
7
83
🧵 I review recent results in two companion papers on market impact. The core contributions: 1) explicit and tractable optimal trading rules with nonlinear impact and decay, and 2) exact "costs" of suboptimal trading due to incorrect expected market impact under an impact model.
1
7
81
@quantymacro
Good question. People usually answer that multicollinearity doesn't matter for random forests. That's wrong. Let your features be column vectors. Random forests partitions the column space into orthogonal subspaces. Collinear data will partially "slip through" the hyperplanes.
2
3
79
"Zero probability" -> infinitely many events in the sample space, they all have equal likelihood of occurring: 1/infinity, which we define as "0".
"Impossible" -> the event "2" is not in the sample space [0, 1], it cannot occur.
Most paradoxes evaporate with good elucidation.
What's your favorite paradox? One of mine is that "zero probability" doesn't mean "impossible". Proof - sample from a Uniform[0,1] distribution! All points have a zero probability, yet you have a sample. But is this really a paradox?
42
7
123
6
1
74
There's no reason for this. We've had very good impact/slippage models with strong empirical backing supported by real-world trade data, at sizes of up to ~3% ADV per day, since 2005. Just use Almgren-Chriss. Linear in vol, square root in participation rate. It works.
5
2
72
I've revised my article following on thoughtful comments from the community. In particular, at their urging some mathematical claims have been made more precise and justified, and some approaches I included for parity with Barra have been critiqued:
2
1
70
@stevehouf
No. The best way to invest in a characteristic is the factor mimicking portfolio specifically constructed with unit exposure to it. Make a well-specified definition of growth, collect data for it on a suitable universe, estimate factor scores, and trade the optimized portfolio.
6
3
68
Note that "good enough" still means a mastery of at least most of undergraduate math, especially analysis, (abstract) linear algebra and probability theory. It also (usually) means the ability to write software that is well designed enough to be maintainable and reliable.
2
1
64
@__paleologo
Okay, I read it. I'll summarize it and provide some commentary. In brief I think it's a useful and credible paper, with specific empirical results, and I'd explore it for further research. But it's not groundbreaking.
4
7
63
@macrocephalopod
@nope_its_lily
Don't understand the causes? Have you never watched people dance and felt the urge to join them? Have you never known love, then felt it slowly fade away until nothing remained?
Then you understand momentum and mean reversion.
5
6
64