Discussion of article "Regression Analysis of the Influence of Macroeconomic Data on Currency Prices Fluctuation" - page 2

New comment
 
Salavat:

So far no other approaches have been found....

That's where you're so wrong.
 
Okay. I won't argue, I'll rephrase: Among the many alternatives, I propose to pay attention to such a method outlined in the article as a promising one
 
Salavat:
Okay. I won't argue, I'll rephrase: Out of the many alternatives, I suggest that the method outlined in the article should also be looked at as a promising one

The tool you have outlined may prove undoubtedly useful in the following circumstances.

1. In works on the construction of a TC with a large number of independent variables, several dozens or hundreds, the approach you have outlined will be useful for marking some search directions. If these formal directions are paired with substantive reasoning about the influence of the independent variable on the dependent variable and about the mutual influence of the independent variables, that is fine.

2. Your proposed tool as a transitional step to a larger set of tools. The point is that STATISTICS as far as I remember is a very limited toolkit and is not up to date. Having made this first step towards analysing the significance of variables for the target variable, it will be natural to move to a larger set of tools, such as those offered by the caret shell in R.

 
Much more questionable is the source of the data. What release of macroeconomic indicators was used? Often such data are marked not with the date of release, but with the date of the end of the period to which they refer. Therefore, using, for example, a GDP series not of the first release one can easily look six months into the future (not to mention revision of calculation methods and redrawing decades of history, as was recently the case with US GDP).
 

The article is interesting. Thank you. But it is more like a guide to using Statistics. And there are a lot of misconceptions in it.

  1. Forecast (and this word appears in the header Regression equation and final forecast) for 5 days using macroeconomic indicators is nonsense. Macroeconomic indicators are released on a monthly and quarterly basis and then adjusted over several months. In addition, macroeconomic indicators have a lot of noise even on quarterly and annual intervals, not to mention daily intervals.
  2. The article does little to explain how these macroeconomic indicators were chosen. It simply takes the 99 most popular ones that are often mentioned in the news. Almost all of them are unable to predict the market as they are lagging, not leading. Even if you take Factory Orders or Building Permits, although they are considered leading, they are not really leading, because they include constituent indicators that were already announced a couple of months ago and to which the market has already reacted. I'll tell you a secret for free: look for different Consumption indicators that you don't have in your data list. They are truly leading. As Consumption went down, Factory Orders and Building Permits went down, and with them GDP and the market. So it all starts with Consumption. This is even well described in Marx's Capital.
  3. Regression analysis is even very applicable to economic models. faa1947 uses some dogmas instead of getting into the essence of mathematical method. The problem is not in the method, but in the data and their preparation. The data should be stationary, no matter whether it is a regression or a "controlled process". Economic data are non-stationary in their original form. But they can be easily transformed into stationary data, for example by differentiation and normalisation.
  4. The problem of regression is that any other series, even completely unrelated to the modelled one, can be inserted into any modelled series, if such input series are chosen sufficiently many. For example, the fluctuations of air temperature in Alaska, data on air pollution in Los Angeles, etc. can be successfully incorporated into forex prices. The regression error can be brought to zero by including such "extraneous data". The accuracy of the forecast will also be zero. Therefore, you need to know how to choose the right data, how much and with what delay.
  5. Knowing the mechanism of how companies function is also useful. For example, everyone likes to quote Unemployment Rate. They consider this rate to be a barometer of the economy. But in reality once the Unemployment Rate has gone up, it is too late to react as the economy is already in decline and has been for a long time. The problem with the UR is that it includes everyone over the age of 16. There are dozens of different unemployment rates for different segments of the population and different occupations. Here's a question for the backfill: if companies see a decline in demand for their products, who do they fire first? And the second question is: which companies are the first to feel a drop in demand?
 
gpwr:

The article is interesting. Thank you. But it is more like a guide to using Statistics. And there are a lot of misconceptions in it.

  1. Forecast (and this word appears in the header Regression equation and final forecast) for 5 days using macroeconomic indicators is nonsense. Macroeconomic indicators are released on a monthly and quarterly basis and then adjusted over several months. In addition, macroeconomic indicators have a lot of noise even on quarterly and annual intervals, not to mention daily intervals.
  2. The article does little to explain how these macroeconomic indicators were chosen. It simply takes the 99 most popular ones that are often mentioned in the news. Almost all of them are unable to predict the market as they are lagging, not leading. Even if you take Factory Orders or Building Permits, although they are considered leading, they are not really leading, because they include constituent indicators that were already announced a couple of months ago and to which the market has already reacted. I'll tell you a secret for free: look for different Consumption indicators that you don't have in your data list. They are truly leading. As Consumption went down, Factory Orders and Building Permits went down, and with them GDP and the market. So it all starts with Consumption. This is even well described in Marx's Capital.
  3. Regression analysis is even very applicable to economic models. faa1947 uses some dogmas instead of getting into the essence of mathematical method. The problem is not in the method, but in the data and their preparation. The data should be stationary, no matter whether it is a regression or a "controlled process". Economic data are non-stationary in their original form. But they can be easily transformed into stationary data, for example by differentiation and normalisation.
  4. The problem of regression is that any other series, even completely unrelated to the modelled one, can be inserted into any modelled series, if such input series are chosen sufficiently many. For example, the fluctuations of air temperature in Alaska, data on air pollution in Los Angeles, etc. can be successfully incorporated into forex prices. The regression error can be brought to zero by including such "extraneous data". The accuracy of the forecast will also be zero. Therefore, you need to know how to choose the right data, how much and with what delay.
  5. Knowing the mechanism of how companies function is also useful. For example, everyone likes to quote Unemployment Rate. They consider this rate to be a barometer of the economy. But in fact once the Unemployment Rate has gone up, it's too late to react as the economy is already in decline and has been for a long time. The problem with the UR is that it includes everyone over the age of 16. There are dozens of different unemployment rates for different segments of the population and different occupations. Here's a question for the backfill: if companies see a decline in demand for their products, who do they fire first? And a second question: which companies are the first to feel a drop in demand?

I have not fully stated my point of view and I will try to make up for it briefly below, maybe repeating myself.

1. The problem of selecting initial data - predictors - is fundamental, poorly formalised and tends more towards art than science. Let's not forget one of the basic postulates of statistics: "Garbage in - rubbish out".

2. As I wrote above, more than half of the time when building a model is spent on selection and justification of the list and type of initial data. Moreover, it is the meaningful interpretation of the initial data, not their formal, statistical characteristics, that is of leading importance. Matapparatus is an auxiliary means for meaningful selection of initial data.

3. According to the literature, I distinguish between two types of forecast: one forecast, the other prediction.

4. Forecast: we take history and extrapolate it into the future a few steps ahead. Classic: we take the mach and extrapolate it forward. The main problem is that the error accumulates as the number of steps forward increases, since the next value is based on the previous one

5. Prediction: we get the current set of input data and predict the future without any prehistory. The previous value is not used, i.e. if we predict +5, we don't need the previous 4 values for that, unlike prediction.

5. In addition to the initial data it is very important WHAT WE PredICT. When trading, we have two types of orders (with options) to buy and sell. For some reason we predict the future price value by regressions and conclude "buy and sell" from this price value. And if we take into account the error and build the decision "buy-sell" taking into account the confidence interval, it quickly becomes clear that it is impossible to make a decision. It follows from this: REGRESSIONAL MODELS PROGNOSING SIGNIFICANCE - UNREPRESENTABLE.

6. One should predict the direction of the trend, which coincides with the orders of the trading systems. This is done by classification models that are able to predict values: "long-short" or "long-side-short", or something else qualitative, but not quantitative, such as: the future value of the pair = 1.3500.

7. For building classification models, this article can be very useful.

 
gpwr:

The article is interesting. Thank you. But it is more like a guide to using Statistics. And there are a lot of misconceptions in it.

Welcome back to the site. ;-) Hopefully for a long time. There will finally be something useful and informative to read.
 

Here is an example of my regression model predicting the S&P500. The black line is the historical index, the solid horizontal blue line is the quarterly average, and the dotted line is the predictions. The predictions are in quarters. The accuracy of predictions is not so great, but it is enough to predict the character of movements: down, up, flat. The model predicts that the market will go down in the remaining 3.5 months of this year, or in the best case it will be flat. I use these predictions only to get out of the market in time.

 

Question or request to the author - in the obtained model, please go to the Advanced tab in the results and click on the Partial Correlations button.

Post here, if not difficult, the value of PC coefficients for each of the model factors.

 
Demi:

Question or request to the author - in the obtained model, please go to the Advanced tab in the results and click on the Partial Correlations button.

Post here, if not difficult, the value of PC coefficients for each of the model factors.

all operations you can do yourself, the file with prepared data is attached to the article in the archive calendar_2010-2011_usd_out. zip .


 b*in Partial Cor. Semipart Cor. Tolerance R-square t(357) p-value
price change for 1 day 0.3500049 0.3883344 0.3045530 0.7571426 0.2428574 7.9622506 0.0000000
price change for 10 days 0.3271369 0.3623699 0.2809839 0.7377407 0.2622593 7.3460567 0.0000000
USD Existing Home Sales (MoM) 0.4499016 0.2746573 0.2064353 0.2105391 0.7894609 5.3970552 0.0000001
USD MBA Mortgage Applications -0.1070692 -0.1315908 -0.0959356 0.8028431 0.1971569 -2.5081463 0.0125795
USD Employment Cost Index 1.1924082 0.3459071 0.2664363 0.0499272 0.9500728 6.9657258 0.0000000
USD NAPM-Milwaukee 0.4918679 0.3281278 0.2510388 0.2604862 0.7395138 6.5631723 0.0000000
USD Existing Home Sales -0.6131716 -0.3510242 -0.2709271 0.1952275 0.8047725 -7.0831333 0.0000000
USD Unemployment Rate -0.2303595 -0.1174173 -0.0854492 0.1375953 0.8624047 -2.2339883 0.0261025
USD ISM Manufacturing 0.4683029 0.2807841 0.2114298 0.2038349 0.7961651 5.5276311 0.0000001
USD Capital Goods Orders Non defence Excluding Air -1.0522008 -0.3326872 -0.2549579 0.0587137 0.9412863 -6.6656324 0.0000000
USD Durables Ex Transportation 1.0195595 0.3332286 0.2554246 0.0627625 0.9372375 6.6778344 0.0000000
USD House Price Purchase Index (QoQ) -0.6493641 -0.3164098 -0.2410558 0.1378030 0.8621970 -6.3021765 0.0000000
USD Chicago Purchasing Manager -0.7364752 -0.2854029 -0.2152134 0.0853930 0.9146070 -5.6265502 0.0000000
USD Personal Consumption Expenditure Core (YoY) -0.5430761 -0.2067342 -0.1527068 0.0790670 0.9209330 -3.9923747 0.0000794
1234
New comment