Hi All,
During this week I want to have a discussion with you about Artificial Neural Network and Time Series Modeling Based Approach to Forecasting the Exchange Rate in a Multivariate Framework. It focuses at improving the predictive performance of these approaches in time series forecasting.
Source: https://arxiv.org/ftp/arxiv/papers/1607/1607.02093.pdf
1. What do you think about accuracy of presented methods
2. How would you approach feature engineering problem for time series forecasting
3. How would you frame time series forecasting problem
Ewelina, no, thanks for your article. At the beginning I would like to ask you what is wrong with you? In my opinion the PhD level students should count to one hundred at least. Your article is almost thirty two pages long!!! Do you know it is so long!!! or you posted it without reading? BTW People please remember t is an English course and articles which you post should be friendly for all of us not only for moderators! If moderators post very long articles (like Ewelina) from new science fields, it is impossible to read such articles carefully and reply with sense. It seemed to me we are adults and we should treat each other with respect. If somebody posts very long and difficult articles it is obvious that majority of students must be cheaters.
ReplyDeleteBack to the topic and your first question I reckon accuracy of presented methods is enough to some activities but is not enough to other. As for your second question I don’t know the answer unfortunately. In my opinion only a seer can predict the future. Your last question is too personal to answer.
At the beginning I am expecting respect from you and I would not accept rude comments like "what is wrong with you". You are grown adult on PhD studies so such behavior should not be acceptable.
DeleteSecond:
I have read this article because this is field of my scientific studies.
Additionally I agree, this my be considered as too long (there was longer articles in the past), so we can address this issue on the forum but without pointing fingers like you have done.
I am fallowing requirements of Dr. Świerk
Please refer to Dr. Świerk mail
"Dear Students,
Please find attached your grades for Week 1. Week 3 finishes today at 11:55 p.m.
Do remember to read articles before presenting comments/opinions.
Secondly, you are to present scientific articles, not the ones from popular magazines or TED presentations (you can add them as an additional source though).
Regards
Małgorzata Świerk"
At the beginning I agree with Ewelina that regardless of our personal opinions this is an English course on PhD level and we suppose to follow the rules of this studies. We suppose to motivate ourselves and constructive criticism. We are not forced to answer all of the published articles.
ReplyDeleteAs an answer to Ewelina's questions.
1. What do you think about accuracy of presented methods.
In the chapter 4.1 conclusions are straight forward. It can be clearly seen that actual exchange rate and predicted exchange rate are very close for training, validation, test and entire data set. Almost a linear trend can be observed between the actual
and predicted exchange rate which justifies the efficacy of the model. All graph shows the same. So I can assume that model created by the authors is very accurate.
I have to add that it is not my field of interests.
2. How would you approach feature engineering problem for time series forecasting
I have no experience with any of this field of science. But the approach is quite obvious because lots of predictions are based on the similar experience. First data set, then model and at the final step, compare reference data with model data. Comparison says something about the accuracy of the created model.
3. How would you frame time series forecasting problem
This is like a reading tea leaves :) If we could do this with 100% accuracy we could predict the stock exchange rate market behaviours. As far as I know it is not happening yet :)
Thank you Rafał,
DeleteCurrently I am framing this problem as supervised learning by providing a set of time series quants for my LSTM network, but I was wondering maybe there is better approach for this
Ewelina, although the form may be unfortunate, I totally agree with the sense of ZC's remarks. This is an extremely technical and very long article, which is a pain to read for those less familiar. We should respect our time and efforts. Just please, consider this without taking it too personal.
ReplyDeleteAs for your questions:
1. The article clearly shows that nowadays very often variations of Artificial Neural Networks algorithm beats every others, even those specific ones that ruled the field for the past years. I'm waiting for the moment that excitement about ANN and, what follows, their dynamic development will reach it's peak. I'm curious what is gonna happen next :)
2. I would probably try LSTM (Long-Short Term Memory recursive neural nets). This is also a modification of ANNs and it's very popular eg. in the field of speech recognition. There is a nice article about implementation of LSTMs: http://machinelearningmastery.com/time-series-prediction-lstm-recurrent-neural-networks-python-keras/
3. I think it's very limited so that when it works I always consider it kind of magic :)
This comment has been removed by the author.
DeleteThank you Ola for your thoughts.
DeleteI have came across article proposed by you some few months ago and it is in fact nice starting point for LSTM implementation
Basing on examples, as well as the conclusions provided by the authors of the article the presented methods seem to be very accurate and the results obtained are very satistfying in terms of predictive efficiency. As per questions 2 and 3, unfortunately I am not an expert in this field and therefore I cannot bring any added value to the discussion, but the topic is very interesting and I will gladly deepen my understanding of the subject by doing some reading on the matter.
ReplyDeleteThank you,
DeleteYes indeed time series forecasting is very interesting for me and can be applied not on for economic data estimation but also many other disciplines such fraud detection, cyber security, utilities and many others problems that are changing over time.
1. What do you think about the accuracy of the presented methods,
ReplyDeleteFrom the article it follows that the results obtained, correspond to reality, so they method are very accurate.
2. How would you approach a feature engineering problem for time series forecasting
3. How would you frame time series forecasting problem
This is not my heritage and research area, without a broader review of literature, I am unable to answer the last two questions.
Thank you Maciej for your time
DeleteThe presented methods are applied to the difficult problem of exchange rate forecasting. Two general approaches to combining neural networks are proposed and examined in predicting the exchange rate. As far as I know the basic ensemble approach created with non-varying network architectures trained using different initial random weights is not effective in improving the accuracy of prediction while ensemble models consisting of different neural network structures can consistently outperform predictions of the single "best" network. Results also show that neural ensembles based on different partitions of the data are more effective than those developed with the full training data in out-of-sample forecasting. Moreover, reducing correlation among forecasts made by the ensemble members by utilizing data partitioning techniques is the key to success for the neural ensemble models. Although methods presented in article show considerable advantages over the traditional KTB approach, they do not have significant improvement compared to the widely used random walk model in exchange rate forecasting.
ReplyDeleteSource: Time series forecasting with neural network ensembles: An application for exchange rate prediction. Available from: https://www.researchgate.net/publication/228492360_Time_series_forecasting_with_neural_network_ensembles_An_application_for_exchange_rate_prediction [accessed May 14, 2017].
Thank you Marcin,
DeleteGreat thoughts about ensembles.
Don’t take this personal but it is true that it is hard to read such complicated and long articles with very scientific data and give constructive opinions.
ReplyDelete1. What do you think about accuracy of presented methods
I am not an expert or person interested in this subject so I can only say what is my impression. It seams that this method works, the results are promising. Conclusions can be
drawn that all the four models have been quite effective for forecasting exchange rate.
2. How would you approach feature engineering problem for time series forecasting
I have absolutely no idea. In my opinion this article exceeds our competences because it is to complicated and to answer this I would have to be expert in this field and spend endless hours on learning this area of science. The technologies used are known for good results. Therefore, I assume that also with this problem, this is a very good solution and I don’t know any other.
3. How would you frame time series forecasting problem
I think that 100% accuracy in probably not possible to obtain. There are also external factors - you can’t predict wars, business collapse, introduction of new companies on the stock exchange or - as we observed lately - countries leaving European Union which changed rates as well.
accuracy of 100% of course is not possible to achieve because this experiment is hard to model by definition. In case where there is this much external factors hard to forecast hole problem is rather challenging
Delete1. What do you think about accuracy of presented methods
ReplyDeleteThe article shows that the methods are working ans the results are satisfactory. I am not able to express my personal opinion, because I have very limited knowledge in presented area.
2. How would you approach feature engineering problem for time series forecasting
I think I am not able to propose any approach.
3. How would you frame time series forecasting problem
Sorry, but time series forecasting problem is too far from my interests and research area, so I am not able to write anything valuable in this subject
Thank you Emilia
Delete1. What do you think about accuracy of presented methods
ReplyDeleteI can relay just on what is written in this article.
2. How would you approach feature engineering problem for time series forecasting
If I have to I would try some stat of the art deep learning methods :)
3. . How would you frame time series forecasting problem
The topic is interesting even though is not my field of research. I've heard about applications of deep learning to time series forecasting (like lstm etc) but I'm not an expert in it.
Thank you Katarzyna for your thoughts
DeleteAd. 1
ReplyDeleteIt seems to be very accurate. To examine the accuracy properly I should do my own research. Models have one problem – exceptions :-) As long as they are not found, we have a rule. And the question is, when to stop looking for “apostasies”.
Ad. 2
I believe everything happens in time cycles and the elements of those cycles are repeatable. That concerns also features engineering, you just have to describe the cycle and estimate the life time of its elements. Next you have to go through as much experiences as you decide to lead.
Ad. 3
Accordingly to my approach mentioned in point 2 above, continuing: all depends on the proper preparing the model and estimating number of experimental data needed. If you do it well, I strongly believe you can achieve 100% of success in predicting.
Thank a lot for your comment
DeleteAt this point with my studies I am incorporating about 60-70 features form LSTM solution, which are mix od previous data and some technical analysis data
Thank you Ewelina for interesting article and I regret that you must face this difficult problem. I used to work with Time series neural network on my second or third semester and us I remember mentioned neural networks are pretty accurate. I try don't go to deep with this algorithm studies because this field is relly difficult. Finally I am finding really fun when somebody is asking me if this types of network can exchange human in their work and this is funny task for managers to tutor younger and less experienced people.
ReplyDeleteYes this is true,
DeleteThanks for your comment
This comment has been removed by the author.
ReplyDeleteBased on this article, charts and tables, I don’t have any reason to reject the hypothesis of the high accuracy of the presented methods. After a brief analysis of the dependency I focused on formula twelve, I see something in recursive style but unfortunately I do not see the boundary conditions. I do not know if I interpret correctly?
ReplyDeleteThank for your comment
DeleteThis article is quite interesting for me for a number of reasons.
ReplyDeleteFirst of all, the topic - forecasting exchange rates.
In my personal opinion, apart from deployed forecasting method it seems impossible to forecast exchange rates accurately. Exchange rates are dependent on high volatility and sensitive to any unexpected occurrence (like transition from fixed to floating FX). It seems that the state-of-the-art models ((G)ARCH/AR(I)MA) were not sufficient to predict FX rates mainly because of neglecting fundamental analysis. The winning method (NARX) takes into consideration external and past events contrary to the AR(I)MA and GARCH models. AR(I)MA and GARCH are models for a single time series and the whole analysis is based on the past events of this particular exchange rate.
I think that idea of introduction of MLFNN and NARX models may become a milestone in time series analysis of high volatility explained variables.
Thank you for your comment.
DeleteMLFNN have some potential within this field, but in my opinion ANN that are implementing back propagation through time like LSTM can even achieve better performance
1. What do you think about accuracy of presented methods
ReplyDeleteIt's hard to argue with the authors of the article. I'm afraid I don't have enough knowledge to do that.
2. How would you approach feature engineering problem for time series forecasting
To be honest, this topic is far from being close to me. After reading linked article I'm in the same place I was before. So, if I want to give clear and honest answer, I can only say I've no idea. Apart from copying authors approach.
3. How would you frame time series forecasting problem
Next question please :)
Thank you for your comment Mateusz
Delete1. What do you think about accuracy of presented methods
ReplyDeleteI haven’t found any summary of accuracy in the article. In “concluding Remarks” there is only a statement that used sets of techniques generated useful and efficient predictions of the exchange rate. In my opinion it leaves a lot to be desired.
2. How would you approach feature engineering problem for time series forecasting
It’s not my cup of tea so it is hard for me to give the right answer. To my mind forecasting of anything is burdened with a mistake. Of course it depends on the time (as a variable) however it will be always a kind of forecasting so we don’t know what will really happen.
3. How would you frame time series forecasting problem
I really don’t know. It’s a difficult topic for me. Moreover I do not like forecasting issue because as I mentioned above it is a variable issue.
Thank your for your thought Damian
Delete