Hacker News Reader

Comment by goodside | Hacker News Reader

This is pure nonsense. This isn’t even the right way to begin thinking about this as a forecasting task — the target series should be log-normal returns, not raw asset price. The performance of this model is laughably bad, which is probably why he spends zero time evaluating its effectiveness. You could trivially get better forecasts than this by naively repeating the last-observed price.

This isn’t ML. It’s cargo-cult performance of words and ideas that ML people use.

View on HN

Even more so since test data can't be from the same time range: i.e. for time series you need to split train/test by date, not randomly, otherwise your model just memorizes the series.

The article splits in time, not randomly.

It's standard practice to validate forecasts on non-randomized test/validation splits of the same time series, since this simulates the conditions where the model will be deployed in reality: It will know everything there is to know about the past, and it will know nothing about the future.

See Hyndman's fpp2 — https://otexts.org/fpp2/accuracy.html

Also, his description of rolling window validation: https://robjhyndman.com/hyndsight/rolling-forecasts/

As someone outside the tech sphere on either coast, that's all ML seems to be.

What I've seen from companies marketing to Higher Education is - we have a lot of data, you set arbitrary flags to the data that you believe indicate 'x' (or even better, they have pre-built data expectations) and you will get 'y' outcome.

And none of it is actually based on anything real. It's all anecdotal applied to extreme amounts of actual data.

And when I read about ML on here, it seems to confirm my experiences.

It's so much worse to see ML applied to problems that could surely be studied and understood mechanistically, but people are sold on using ML instead in deference to buzzwords alone. The result is that they may get some model of a phenomenon, but they'll never learn a goddamn thing about why it works. What good is that?

It depends on what your goal is; sometimes it’s the destination, sometimes it’s the journey.

The problem isn’t that ML has gotten worse. It’s just as rigorous and far more powerful than it ever was. The problem is ML is hard, it hasn’t gotten orders-of-magnitude easier to understand, and there’s enormous incentive now to pass off amateur understanding as complete. The real ML still happens — it’s just drowned out.

I think people are saying that that's something they're not happy with. The big methods in AI, like backprop,

1) work a LOT better for specific problems than statistics or statistical learning ever has (and at this point, I think we can safely say: ever will)

2) a lot of methods either can't be explained, or outright shouldn't work, according to statistical theory.

The use of statistics in machine learning is limited to evaluating performance and individual element performance (and even that is tenuous at best in many cases). If you ask, say, why would an autoencoder, with an LSTM on it's compressed representation and Q-learning evaluation have somewhat decent performance on half the computer games humans ever designed ? Statistics will not be useful in formulating an answer.

If you ask extremely valid questions, like "why would an LSTM predict anything ?". Statistics draws a blank. There is no good reason to assume an LSTM will ever converge (and on a truly random dataset, it won't, whereas statistical methods will still allow you to say something).

I think there's 2 reasons for this

1) the "upper limit" of complexity a human can understand in a statistical model is lower than the upper limit a neural network can "understand". In statistics the human understanding is critical to getting to a valid model, in machine learning ... it is not. Meaning machine learning can learn relationships a human mind cannot.

2) There must be some fundamental property of the world we live in that matches neural network architecture. In order for backprop to work on real-world problems, it has to be the case that almost all real world phenomena are continuous, both "raw" and in the frequency domain. If this wasn't the case, machine learning would never be able to learn anything.

can't be explained, or outright shouldn't work, according to statistical theory

When people invented a steam powered engine, some other people had probably said: "Modern physics can't explain how it works. It shouldn't work. It's too complicated". Then a few decades later physicists discover laws of thermodynamics.

> that's all ML seems to be.

You have to be very selective about what you consider "ML" to come to that conclusion. There has been a constant parade of incredible, mind-blowing results out of ML over the past decade, advancing the state of the art by leaps and bounds both in research and in real applications.

Do you not remember how terrible speech recognition and speech synthesis were just a few short years ago? Did you not see DeepMind finally crack Go? Check out BigGAN [1]. Try search in Google Photos. See translation getting better every year.

Yes, there are quacks and charlatans and people who are just plain wrong. But ML is real, and it solves real problems that people failed to solve any other way despite decades of concerted effort.

[1] https://medium.com/syncedreview/biggan-a-new-state-of-the-ar...

I concur totally, however think we should add in a caveat that is relevant for this audience: ...only the major consumer tech companies are actually reaping the benefit of applied ML in a way that is profitable.

That's because they have the platforms and applications that people are using at scale. So ML is a force multiplier if you already have a consistent and strong user base for a good product.

If you're trying to get a product or company started, unless you're a pure ML research company like Clarifai (and arguably even then), ML is probably going to cost you more than you gain.

These days, startups founded by people who have worked in big tech companies and have witnessed the power of ML/AI will definitely apply it from day-1.

You don't have to invent or implement the algorithm yourself. You can use AI/ML intelligent/cognitive services provided by the big-3 cloud companies to reduce your starting cost significantly.

To have positive ROI, your problem complexity should have crossed the threshold where common-sense traditional solutions don't work any more.

If you do embark on ML research yourself, then be sure to walk the path from simpler models to complex ones while carefully establishing performance metrics.

Can you clarify why "the target series should be log-normal returns" is important or provide a pointer for more information?

Investment markets operate on relative gain, not absolute gain. E.g.: if you invest in a stock and it gains $5, this would be a great return for a $1 stock but a poor one for a $1000 stock, so the absolute gain doesn't mean anything on its own. A 5% return always means that you've gained 5% on your investment.

Would it be the same (valid) with percentage returns?

Less so. Log-normal returns are better because they have the property that a summation of log-normal returns over contiguous intervals is equal to the log-normal returns of the combined interval. In other words: Losing 5% and then gaining 5% doesn’t put you back at exactly 100%, and log-normal fixes that.

In the extreme, two successive trades, where the first gains 110% and the second loses 100%, “average” out to a 5% return. However, you don’t want to make that pair of trades.

Ah, that makes sense. Many thanks.

I think their point is that predicting that "the stock will go up by 1$" or "the stock will go down by 1$" is worse than "the stock will go up by 0.05%" and "the stock will go down by 0.05%" because of this little paradox:

50$ increase from 50$ is 100% increase 50$ decrease from 100$ is 50% decrease

e.g., if the model finds 50$ increase/decreases, that actually corresponds to very different wealth changes

Well, just at a minimum... There's the fact that the majority of large gains and losses exist over single-day frames. That is, guessing "right" or "wrong" on movement means little when one slip on the wrong day will decimate your returns.

Here you go:

https://financetrain.com/why-lognormal-distribution-is-used-...

Thank you.

https://en.wikipedia.org/wiki/Stationary_process

Welcome to ML in 2018.

http://scikit-learn.org/stable/modules/generated/sklearn.pre... Should normalize the data and raw prices won’t be used and instead the normalized coefficients are effectively a percentage

It's the AI winter we're all fearing!

Clearly one LSTM didn't work. Lets try FOUR!

- edit: miscounted number of LSTMs

Reply on news.ycombinator.com