Market prediction based on macroeconomic indicators - page 43
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Well, here's a clear thought by Hazin.
For 2010, total financial instruments on NYCE NASDAQ = $17.796 + $12.659 trillion. And GDP is half that. And finance's share of GDP is laughable. How can this be?
Everything about statistics in the US has to be done with extreme caution. One has to delve into the methodology of calculating GDP... Is this necessary?
GDP counts the amount of financial services rendered in a year, not the nominal capitalisation of the national stock market
Well, you see, you've made it two. So the brokerage margin is taken into account. It's less than ours - a ridiculous amount.
And the rest of the money? You buy the trousers, it's a sale, but you buy a stock, a futures, what's that? You have to count the turnover of accounts...
Well, you see, you've made it two. So the brokerage margin is taken into account. They have less than ours - a ridiculous amount.
And the rest of the money? You buy the trousers, it's a realization, but you buy a stock, a futures, what's that? You have to calculate the turnover of the accounts...
The whole market economy ....
You have to count in kind, in bigmouths. And the monetary value, and even in different currencies is workable only with stable prices and stable rates. And today? Everyone is stuck with the ruble... And look at the charts of currency pairs! Or look at the charts of the dollar index. What prices can we talk about?
What are we comparing it to? Part of US GDP for electronics with their real, physical production in China? Air with real iron?
As I wrote before, I don't choose the predictors, but the code by prediction error. I just look at the final model and check what the indicators are. For example, one indicator is the start of house building:
https://research.stlouisfed.org/fred2/series/HOUST1F
Looking at this chart you can see a trend of declining house building starts before recessions. Which indicators are chosen by the code depends quite a bit on the data conversion method. I quite accept that futures have an impact on GDP. If there is a time series reflecting futures, show me the link. All my predictors are taken from the Fed's FRED2 database. There are economic and financial indicators:
https://research.stlouisfed.org/fred2/categories
The problem with these indicators is that there are about 300 thousand of them. Lots of regional and international data. I had to discard "unnecessary" data manually to avoid loading the code. We got about 10 thousand indicators. But not all of these indicators have been released by the government for a long time: some of them have been released since 1800, others started to be released in the last 10 years. My code only considers the indicators which have been produced since 1960 to have enough data to build the model, i.e. 55 years of history or 220 data in each indicator. The 10 thousand selected indicators narrows down to 2 thousand. One could go into the discussion that the market today is significantly different from the market 50 years ago. And I quite agree with that: computers, internet, Chinese influence, US government interference in economic management, and so on. But if I only take data from the last 15 years, I only get 60 values in each indicator, only two recessions, and that, as statisticians know, is not enough to build a model. So I have to go deeper into history, when the economy was different. As a result, other difficulties arise: what described the economy then, does not describe it well today. By the way, I tried shortening history by 15-20 years, but the predictions were much worse.
As I wrote before, I don't choose the predictors, but the code by prediction error. I just look at the final model and check what the indicators are. For example, one indicator is the start of house building:
https://research.stlouisfed.org/fred2/series/HOUST1F
Looking at this chart you can see a trend of declining house building starts before recessions. Which indicators are chosen by the code depends quite a bit on the data conversion method. I quite accept that futures have an impact on GDP. If there is a time series reflecting futures, show me the link. All my predictors are taken from the Fed's FRED2 database. There are economic and financial indicators:
https://research.stlouisfed.org/fred2/categories
The problem with these indicators is that there are about 300 thousand of them. Lots of regional and international data. I had to discard "unnecessary" data manually to avoid loading the code. We got about 10 thousand indicators. But not all of these indicators have been released by the government for a long time: some of them have been released since 1800, others started to be released in the last 10 years. My code only considers the indicators which have been produced since 1960 to have enough data to build the model, i.e. 55 years of history or 220 data in each indicator. The 10 thousand selected indicators narrows down to 2 thousand. One could go into the discussion that the market today is significantly different from the market 50 years ago. And I quite agree with that: computers, internet, Chinese influence, US government interference in economic management, and so on. But if I only take data from the last 15 years, I only get 60 values in each indicator, only two recessions, and that, as statisticians know, is not enough to build a model. So I have to go deeper into history, when the economy was different. As a result, other difficulties arise: what described the economy then, does not describe it well today. By the way, I tried shortening history by 15-20 years, but the predictions were much worse.
As an old acquaintance of mine who lives in Canada says: your number eight, you will be asked afterwards.
It's important to know your place in this queue.
It's somehow demeaning and insulting to ourselves. Is there no sense of dignity? You won't get very far with such an attitude.
I am not an expert in economics and I don't know how wise number 1 is, but in my field of science I can assure you that university professors know far less than people working for companies, practitioners. I suppose it's the same in economics: only a couple of experts who are capable of new theories, and the rest are, as the Americans say, polishing the apple. Do you think that the fed banks employ luminaries who know how to predict the economy? What about 2008? Bernanke refused to lower rates until September 2007, 3 months before the recession officially began. And what about the Long-Term Capital Management hedge fund that went bankrupt in the late 1990s and was bailed out by buyout banks and the government? This fund was headed by two Nobel laureates Scholes and Merton.Scholes, as you may know, is one of the authors ofthe Black-Scholesfinancialmodel(option pricing model) for which he received his Nobel Prize. Why is it that rich investors (and those were the members of LTCM) either make good money or lose money but always get it back, from the government at most, while other investors lose money on the stock market and that's it, no one bails them out.
https://en.wikipedia.org/wiki/Long-Term_Capital_Management
Honestly, I haven't read it myself, but has anyone read Didier Sornetto?
I will outline my experience with input data transformation. There are several ways of transforming the data outlined in articles on economic modelling:
1. Difference: x[i] - x[i-1]. Applicable if input x[] has constant variance. I have about 2 thousand predictors with histories dating back to 1960. To see how their variance varies with time, I calculated the difference x[i] - x[i-1], squared it, then averaged it using Hodrick-Prescott filter with lambda 1e7 and took the root to see the variance as a function of time. Then I divided the variance at the end of history (Q4 2015) by the variance at the beginning of history (Q1 1960) for each input variable and made a histogram:
Many inputs have more or less constant variance (the ratio of variance at the beginning to the end of the story is about 1). But there are also a lot of inputs with a variance ratio of 3 or more. The variance of GDP increases about 4 times from 1960 to today. Since it is not possible to construct a GDP model with inputs whose variance does not change, transforming the inputs by the variance is insufficient.
2. Momentum: x[i]/x[i-1] - 1 or log(x[i]/x[i-1]). Automatically normalizes inputs with different variance, but only works if all data are positive. The formula x[i]/x[i-1] - 1 = (x[i] - x[i-1])/x[i-1] can be thought of as calculating growth in %, i.e. x[i] - x[i-1] as a percentage of x[i-1]. If x[i-1] is zero, this formula makes no sense and gives an infinite value. When x[i-1] is negative, this formula does not make sense either. Approximately 15% of economic indicators have both positive and negative values. You can try to use momentum for positive series and variance for negative series, in the hope that negative series have approximately constant variance. Unfortunately, there are some economic indicators that have positive and negative values and the variance grows strongly over time. For example:
3. The variance normalized by the variance is (x[i] - x[i-1])/StdDev[i]. In my experience this is the best and most versatile transformation suitable for all kinds of data. There are two serious problems here: (1) how to correctly calculate the time dependent variance of StdDev, and (2) how to convert the prediction back to the original series form if the future variance is unknown.
I would divide all modern economic mathematics into two parts
The division would seem to be wrong, as it is impossible to predict the future without analysing the past.
In practice, however, I have found that this is not the case. A distinction exists, and a fundamental one at that.
1. There is analysis in itself. We analyze unemployment and look for the factors that influenced it in the past.
2. And there is another analysis. Initially we try to predict unemployment and look for the factors which influenced this unemployment in the future.
In the first case, if we want to predict the future, we extrapolate the results of our analysis. Here, we encounter a situation in which the difference between the extrapolated value and the current value falls within the confidence interval, which means that the best predictor based on the extrapolation is the current value!
In the second case, we are not interested in the previous value. We calculate a new, future value (trend) based on historical data, rather than continuing the past into the future. In this case, when new data arrives, the model makes a prediction based on knowledge of past situations, which are not necessarily previous, but were in the past.
That is, extrapolation should be rigorously distinguished from prediction.
The seemingly subtle differences entail very serious consequences.
1. The target variable itself. It turns out to be far from an idle question. It's impossible to do without thorough analysis of the target variable's properties, especially taking into account item 2.
Selection of predictors that are relevant to the target variable. Selection of predictors that have predictive potential for the target variable for its properties. For example, the target variable: growth-decline. We need predictors that are relevant to the growth-decline of the target variable, but are not interested in predictors that predict the value of the target variable.
PS.
From experience. In this approach of predicting nominal variables, I have found no effect on the predictive ability of predictors by pre-processing them as described above, and by more radical methods such as transformation into a set of principal components (PCA or others), which have surprising properties for us, and no use.