Aspiring Data Science
327 subscribers
390 photos
10 videos
6 files
1.46K links
Заметки экономиста о программировании, прогнозировании и принятии решений, научном методе познания.
Контакт: @fingoldo

I call myself a data scientist because I know just enough math, economics & programming to be dangerous.
Download Telegram
#finance #algotrading #ml #businessidea

Эрни Чан - бывший сотрудник хедж-фонда, уже много лет как частный трейдер, автор 3 хороших книг и множества лекций. Очень интересно наблюдать эволюцию его взглядов на трейдинг и применимость ML в финансах. Идея об уходе от общих таргетов заслуживает внимания и проверки, тем более что о ней же говорит и ДеПрадо (в его книге эта техника называется "метаразметкой").

https://www.youtube.com/watch?v=nAOjbL5bsjA

Книга Quantitative Trading: How to Build Your Own Algorithmic Trading Business https://www.amazon.com/gp/product/B097QGPVND/

Ещё у него есть активный блог: http://epchan.blogspot.com/
#trading #zhang #finance

Недавно постил лекцию Жихонг Жанга по DL на MOB, вы только посмотрите, насколько плодовит данный учёный!
Мама мия, 37 работ, и многие очень интересны уже по заголовкам. А, или то был Жихао. Все молодцы, в общем )
С нетерпением начту прочтение их совместной работы с Хуйлинь Юанем, не побоюсь этого имени, "Forecasting security's volatility using low-frequency historical data, high-frequency historical data and option-implied volatility".

https://arxiv.org/search/q-fin?searchtype=author&query=Zhang%2C+Z
#trading #finance #lob #dl

Очень интересные открытия: нет следов большой нестационарности (даже для 1 года OOS), 1 модель для всех активов лучше отдельных моделей для каждого актива. Использование моделек с памятью типа LSTM улучшает точность (от 100 до 5000 тиков, к примеру, на 1% - path dependence/long memory effects).

https://www.youtube.com/watch?v=diLtyRg6cl4
#finance

Это просто бриллиант какой-то. Абсолютно бесполезная по содержанию лекция, но докладчик посетил курсы ораторского мастерства. Как будто в театр сходил! На 20й минуте заснул. Советую против бессонницы.

https://www.youtube.com/watch?v=z-HHsUfc1C8
#emh #finance #physics

Внезапно увидел аналогию в финансах, машинном обучении и физике. Есть такая EMH, гипотеза эффективного рынка. А кто угадает, о какой аналогии я подумал в машинном обучении и даже, если разобраться, в астрофизике? Какие там существуют похожие концепции?
#emh #finance #physics #nflt #ccc

Ну ладно, раз уж практически всё отгадали, публикую ответ. Конечно же, это ML-теорема об отсутствии бесплатных завтраков (No Free Lunch theorem) и Принцип космической цензуры (Cosmic Censorship Conjecture) Пенроуза.

Вы только подумайте, просматривается связь физики, машинного обучения и поведенческих финансов. Возможно, физику я несколько притягиваю за уши, судите сами. Я попросил сами знаете кого обозначить параллели.

"The Cosmic Censorship Conjecture (CCC) in physics, which suggests that singularities (places where gravity becomes infinite) are hidden behind event horizons, has intriguing parallels with the Efficient Market Hypothesis (EMH) and the No Free Lunch Theorem (NFLT) in machine learning. All three concepts imply that exploitable extremes or shortcuts are fundamentally "censored" by the natural order of the system, and they share some core ideas:

1. Hidden Extremes and Limited Access

Cosmic Censorship Conjecture: Singularities (places with potentially infinite gravitational forces) are hidden behind an event horizon, so they can’t affect or be exploited by an external observer. This keeps the universe predictable and prevents runaway effects from unbounded gravitational fields.

Efficient Market Hypothesis: Similarly, in EMH, any “infinite gain” opportunity is concealed by the rapid incorporation of information into asset prices. Just as a singularity is hidden from view, a perfect arbitrage opportunity (extreme profit) is obscured by market efficiency, making it virtually unreachable.

No Free Lunch Theorem: In machine learning, NFLT states that no model is universally better across all datasets, implying that there are no "shortcuts" that work everywhere. Every dataset poses new challenges, and there’s no access to a "singularity" of knowledge or performance that guarantees success universally.

2. Equilibrium and Stability

Cosmic Censorship Conjecture: By hiding singularities, CCC maintains stability in spacetime and keeps the equations of general relativity well-behaved, creating a predictable universe.

Efficient Market Hypothesis: EMH maintains market stability by ensuring that all known information is reflected in prices, which keeps the market in a kind of equilibrium. When information is immediately priced in, it prevents drastic, destabilizing price swings due to secret information exploitation.

No Free Lunch Theorem: NFLT implies an equilibrium of algorithmic performance: without tailoring models to specific problems, performance can’t exceed random guessing across all tasks. This prevents the dominance of any one algorithm and promotes stability in ML development.


3. Limits on Observability and Exploitability

Cosmic Censorship Conjecture: In CCC, the "naked" singularities (unshielded by an event horizon) are prohibited, meaning that singularities can't be directly observed or interacted with. This places fundamental limits on what can be known about the interior of black holes and ensures that any infinite values are hidden from observers.

Efficient Market Hypothesis: In EMH, since all information is reflected in prices, potential inefficiencies that traders might exploit are hidden or fleeting. This places a limit on how much information a trader can “observe” to gain a consistent edge.

No Free Lunch Theorem: NFLT in ML implies that there's no one-size-fits-all model, so knowledge about one dataset does not translate to others. The unpredictability of the optimal approach for new data is a kind of “observability limit,” preventing universal exploitability of one algorithm.
#emh #finance #physics #nflt #ccc

4. Self-Correcting Systems

Cosmic Censorship Conjecture: If any unshielded singularity exists, it would theoretically destabilize the fabric of spacetime, causing unpredictable consequences. By “shielding” them, CCC ensures that the universe operates as a self-correcting system.

Efficient Market Hypothesis: EMH similarly implies that markets are self-correcting. If someone finds a profit opportunity, the market rapidly adjusts as participants act on it, neutralizing the opportunity—effectively “censoring” it through price adjustments.

No Free Lunch Theorem: NFLT also implies a self-correcting mechanism: the variability of datasets corrects for any over-reliance on a single model. In essence, no shortcut applies universally, so the diversity of data prevents overfitting to any one approach.

5. Conservation of Predictability

Cosmic Censorship Conjecture: By hiding singularities, CCC ensures that the universe remains predictable on a macroscopic level, with singularities hidden in a way that preserves the general structure of spacetime.

Efficient Market Hypothesis: EMH enforces predictability in financial markets by ensuring that prices reflect all known information, preventing “wild” deviations or exploitable patterns that would otherwise exist in an inefficient market.

No Free Lunch Theorem: NFLT reinforces the conservation of predictability by implying that machine learning performance is fundamentally limited by the variability of data. This lack of universal shortcuts keeps machine learning practice tethered to specific, data-driven methods without general predictability.

In essence, Cosmic Censorship Conjecture, Efficient Market Hypothesis, and No Free Lunch Theorem all maintain system stability, equilibrium, and self-correction by “shielding” the unbounded extremes that would otherwise lead to unpredictability and exploitation. They serve as guardrails in complex systems, placing inherent limits on predictability, observability, and exploitability."