Sociopath Community

AliceInWonderland said: 

LiYang said: 

I've been thinking about machine learning but haven't had the time to put into the research.

 

Forecasting on metal resource spot settlement price: New evidence from the machine learning model

"We found that LSTM-GRU and other models also perform well with strong robustness. Therefore, we believed that the LSTM hybrid model, especially the LSTM-GRU model, is suitable for analyzing the prediction of spot settlement price of metal minerals."

https://www.sciencedirect.com/science/article/abs/pii/S0301420723000685

 

Maybe a simple natural gas economic model with supply and demand inputs. Start simple and see what others have done.

Forecasting doesn't matter as it doesn't inform your actual PNL. 

You can be right about the behavior of a random variable X and still go bust if your pay off function f(x) doesn't align with what your forecasting. 

Moreover, given X in markets is a highly stochastic variable that is fat-tailed its very difficult to accurately forecast given the complexity of that variable. As such, almost all your focus should be towards your payoff function f(x) given its far simpler and doesn't rely on making predictions but instead odds and optimal betting. 

All the stupidity in statistics can really be summarized as the focus on values of functions instead of the functions of values. 

 There are indications that 2023 is going to be a time of recession. I might get them if I must.

Personally I find patterns very suggestive as to what's about to happen. 

I'm not sure what's a good hedge against recessions either, cause people will liquidate their assets after maxing out their credit cards.

AliceInWonderland said: 

LiYang said: 

I've been thinking about machine learning but haven't had the time to put into the research.

 

Forecasting on metal resource spot settlement price: New evidence from the machine learning model

"We found that LSTM-GRU and other models also perform well with strong robustness. Therefore, we believed that the LSTM hybrid model, especially the LSTM-GRU model, is suitable for analyzing the prediction of spot settlement price of metal minerals."

https://www.sciencedirect.com/science/article/abs/pii/S0301420723000685

 

Maybe a simple natural gas economic model with supply and demand inputs. Start simple and see what others have done.

Forecasting doesn't matter as it doesn't inform your actual PNL. 

You can be right about the behavior of a random variable X and still go bust if your pay off function f(x) doesn't align with what your forecasting. 

Moreover, given X in markets is a highly stochastic variable that is fat-tailed its very difficult to accurately forecast given the complexity of that variable. As such, almost all your focus should be towards your payoff function f(x) given its far simpler and doesn't rely on making predictions but instead odds and optimal betting. 

All the stupidity in statistics can really be summarized as the focus on values of functions instead of the functions of values. 

Seems you would want a model for gas supply and then a model for gas demand. Then compare the two. Train your machine learning model on old data sets and back substitute to correct for newer data. Leading input indicators like cold weather would be useful.

I don't understand why you would say forecasting doesn't matter. that's kinda the goal, to forecast the price so you can buy or sell.

But I agree, any randomness will need to be identified and filtered out.

AliceInWonderland said: 

LiYang said: 

I've been thinking about machine learning but haven't had the time to put into the research.

 

Forecasting on metal resource spot settlement price: New evidence from the machine learning model

"We found that LSTM-GRU and other models also perform well with strong robustness. Therefore, we believed that the LSTM hybrid model, especially the LSTM-GRU model, is suitable for analyzing the prediction of spot settlement price of metal minerals."

https://www.sciencedirect.com/science/article/abs/pii/S0301420723000685

 

Maybe a simple natural gas economic model with supply and demand inputs. Start simple and see what others have done.

Forecasting doesn't matter as it doesn't inform your actual PNL. 

You can be right about the behavior of a random variable X and still go bust if your pay off function f(x) doesn't align with what your forecasting. 

Moreover, given X in markets is a highly stochastic variable that is fat-tailed its very difficult to accurately forecast given the complexity of that variable. As such, almost all your focus should be towards your payoff function f(x) given its far simpler and doesn't rely on making predictions but instead odds and optimal betting. 

All the stupidity in statistics can really be summarized as the focus on values of functions instead of the functions of values. 

 um ok thanks i think >.>

ok so if i build the website you math nerds will help me with the machine learning and math stuff? :D

Golden_Eagle said:

There are indications that 2023 is going to be a time of recession. I might get them if I must.

Personally I find patterns very suggestive as to what's about to happen.

I'm not sure what's a good hedge against recessions either, cause people will liquidate their assets after maxing out their credit cards.

 Under great uncertainty not placing a bet is the best decision. 

LiYang said:

I don't understand why you would say forecasting doesn't matter. that's kinda the goal, to forecast the price so you can buy or sell.

But I agree, any randomness will need to be identified and filtered out.

We do not want to filter out randomness, we want to make optimal decisions under highly random conditions. 

One of the fundamental problems with forecasts is that they do filter out randomness which in turn obscures your decision making and leads you to ruin. 

Forecasting errors are necessarily thin-tailed while market events and therefore payoffs are thick-tailed. As such,  forecasts do not allow you to consider events that lead to ruin. Essentially, a highly accurate forecast cannot account for one extremely costly payoff that makes you go bust. Despite that event being rare, it will completely destroy any gains you made prior to that event taking place.   

puppygirl said:

um ok thanks i think >.>

 You're welcome. 

edit :

Source equations for the code are from one of talebs moocs, 

puppygirl said: 

ok so if i build the website you math nerds will help me with the machine learning and math stuff? :D

 Perhaps work on your negotiating skills.

AliceInWonderland said: 

Golden_Eagle said:

There are indications that 2023 is going to be a time of recession. I might get them if I must.

I agree, things could be rough in general in 2023 and beyond. 

 

Even if we don't see a spike in unemployment and other recessionary traits, we can expect securities including crypto to struggle in an environment where interest rates are continuously on the rise and the FED substantially shrinking its balance sheet. 

 

For example, look at how interest rates generally upper bound of price. 

 

 

Notice how the upper bound of price becomes less as R (interest rate) increases from 0, 0.03, and 0.05. 

 

In the above simulations I am assuming price is a geometric brownian motion and creating 500 price paths per a rate. Its worth mentioning that this simulation method is faulty given the method of simulation itself is a random variable, as such you expect each simulation to vary. However, I have created far more robust simulations where I have measured the variance between runs at each rate and it seems true that security prices are are effected by rates. 

 

Personally I find patterns very suggestive as to what's about to happen.

Patterns can be suggestive, and a observable pattern is not necessarily a forecast. For instance, I can measure the probabilistic properties of a basket of securities such as stocks and derive their common properties. Such a methodology can be considered a form of pattern recognition given its whole purpose is to derive properties from a set of observations, but I am not predicting the future or creating a forecast. I am merely trying to verify that security X is equivalent to stochastic system Y by concluding that they both have the same properties. (for the CT folks basically you are establishing an isomorphism between a real world system and a model). 

 

I'm not sure what's a good hedge against recessions either, cause people will liquidate their assets after maxing out their credit cards.

 Under great uncertainty not placing a bet is the best decision. 

 Your chart looks like Ryu's super move.

I'll see if I can find the data I was refering to, it'd be easier if I can find the video explination, it's very interesting. All pattern recognition based on economic performance, and we're reached the top.

Spatial Mind said:

Your chart looks like Ryu's super move.

I'll see if I can find the data I was refering to, it'd be easier if I can find the video explination, it's very interesting. All pattern recognition based on economic performance, and we're reached the top.

Please do. 

A quick google returns only results related to street fighter.  

AliceInWonderland said: 

Spatial Mind said:

Your chart looks like Ryu's super move.

I'll see if I can find the data I was refering to, it'd be easier if I can find the video explination, it's very interesting. All pattern recognition based on economic performance, and we're reached the top.

Please do. 

A quick google returns only results related to street fighter.  

Thomas Kralow presented this yesterday, and it's what I was talking about. It basically shows data pertaining to how it's impossible to go from 3.4% Unemployment to 4.6% without first going into recession. This data goes back to the 40's I believe, which would mean this data has been consistent 11 consecutive times. 

The site has a paywall so skip it if you wish. The video has it the details

https://ecoinometrics.substack.com/p/ecoinometrics-the-federal-reserve

This video should be time stamped from 4:08

.

.

.