Forecast market by photo
- Thread starter Eduard_OCF
- Start date
Maybe sometime.
I tried it several times. Not bad at all, 50%, 66%, 66%, 75% accurate in casual testing on four instruments.
When it forecasts up or down in 10 days, are you saying it'll close up/down on that tenth day; or are you saying that within the next ten days, there'll be a close higher/lower than the most recent close?
How does the app handle charts with more bars than 40-50? Is it OK to capture more than 50 days?
How well does the app recognize the chart? If typical data and artifacts above, below, to the left and right of a chart are present, does that affect the results. I'm not referring to when it asks you to resubmit an image, I'm referring to the app mistaking extra information as being part of the chart (partial lower volume bars/indicators; price data; bar time; name of instrument and other text)
Without knowing your responses yet, here are some of my ideas:
When the app accepts a chart, return an image that somehow highlights/selects/identifies what it believes is that actual chart data; then allow the user to resubmit a capture if the app is in error.
This process will "teach" the user what the app "likes" and doesn't like in captures.
And, or:
Add cropping functionality. This will allow for quicker and easier capturing of the screen, since the "fine tuning" can be done with the cropping tool.
I thought of this because I was trying to capture a particular bar, that I wanted to be the simulated most recent bar. This bar was a historical bar, and had bars to the right of it. So it was a little tricky trying to stay between the bars, while also dealing with my phones attempts to stabilize the image.
During normal use, this may not be an issue, it all depends on whether artifacts to the right of the last bar affect the algo's determination of the bars. And making sure the app "picks" the correct last bar, in my uninformed opinion, would be important.
Here is an example. Would the horizontal last price line interfere? How 'bout the green highlighted price, or the USD icon?
Have you backtested this app/algo, manually or otherwise? How did it do?
Nice work!
When it forecasts up or down in 10 days, are you saying it'll close up/down on that tenth day; or are you saying that within the next ten days, there'll be a close higher/lower than the most recent close?
How does the app handle charts with more bars than 40-50? Is it OK to capture more than 50 days?
How well does the app recognize the chart? If typical data and artifacts above, below, to the left and right of a chart are present, does that affect the results. I'm not referring to when it asks you to resubmit an image, I'm referring to the app mistaking extra information as being part of the chart (partial lower volume bars/indicators; price data; bar time; name of instrument and other text)
Without knowing your responses yet, here are some of my ideas:
When the app accepts a chart, return an image that somehow highlights/selects/identifies what it believes is that actual chart data; then allow the user to resubmit a capture if the app is in error.
This process will "teach" the user what the app "likes" and doesn't like in captures.
And, or:
Add cropping functionality. This will allow for quicker and easier capturing of the screen, since the "fine tuning" can be done with the cropping tool.
I thought of this because I was trying to capture a particular bar, that I wanted to be the simulated most recent bar. This bar was a historical bar, and had bars to the right of it. So it was a little tricky trying to stay between the bars, while also dealing with my phones attempts to stabilize the image.
During normal use, this may not be an issue, it all depends on whether artifacts to the right of the last bar affect the algo's determination of the bars. And making sure the app "picks" the correct last bar, in my uninformed opinion, would be important.
Here is an example. Would the horizontal last price line interfere? How 'bout the green highlighted price, or the USD icon?
Have you backtested this app/algo, manually or otherwise? How did it do?
Nice work!
Dear userque!
First of all I want to thank you for your time. Thanks for testing the app!
I will try to answer all your questions.
At this point, I just took the average price over the next 10 days. At the moment, I consider this a sensible compromise. Although, of course, we can calculate (predict) a variety of values. Obviously, for a professional forecast, we are interested in not only the average, but also the deviations from it, and also correctly apply the threshold (above or below which) we consider the percentage up or down. Which of these to use, I think, depends on who this program is designed for, for beginners or professionals. It is right to give the user a choice of what he needs, what he gets, a professional package and approach or simple.
I do not cut out the extra candlesticks, as I do not add the missing))
In fact, instead, the time series shrinks or expands.
There is another interesting question. But what if we apply it not to daily candlesticks, but to five minutes.
Despite the fact that I consider the time series to be scale invariant, I also believe that for each scale it is necessary to build your own base with a pattern. And in the future I will go this way.
Then your and my question will not arise.
I conducted a huge number of experiments to recognize a wide variety of photos and screenshots. There have already been several implementations. At the moment, I settled on the following approach. I assume that by taking a screenshot the user can accurately and accurately cut out the area of interest. Conversely, a photo may have more errors, so it is subjected to more processing.
In general, I remove vertical and horizontal borders with unnecessary information. Random horizontal lines or indicators are also removed. But it’s better not to allow this.
I assume that a user who wants to get a quality forecast can train to make high-quality photos or screenshots in 5 minutes. It would be a desire.
Great idea!
I thought about it.
Thank!
This feature is usually built into the functionality of the smartphone’s camera.
In my experience, it is usually enough to switch between 16x9 and 4x3 mode, crop is not needed.
About this image. If this were a photo, then all unnecessary information will be deleted, including a horizontal line. If this is a screenshot, then I assume that the user left only the prices on the screenshot, although some of the extra information will still be deleted, but something will remain.
Recommendation for this image and others: Avoid hollow candlesticks. They should be painted in a color different from the background color.
This is not a signal system. These are statistics based on similar situations in the past. I did not train the system in the sense of learning. There is mathematics, there is a strict selection criterion. If you understand the issue of finding similarity in time series, you will understand what I mean.
The only thing I have to take care of is that the historical base be large. Now it is 170 thousand patterns. I think there should be several million. How to get such a base? There are several ways. But this is a completely different conversation ))
Thanks again for testing, feedback and questions.
All the best! Success in trading!
I will be glad if you decide to continue communication on this topic.
First of all I want to thank you for your time. Thanks for testing the app!
I will try to answer all your questions.
At this point, I just took the average price over the next 10 days. At the moment, I consider this a sensible compromise. Although, of course, we can calculate (predict) a variety of values. Obviously, for a professional forecast, we are interested in not only the average, but also the deviations from it, and also correctly apply the threshold (above or below which) we consider the percentage up or down. Which of these to use, I think, depends on who this program is designed for, for beginners or professionals. It is right to give the user a choice of what he needs, what he gets, a professional package and approach or simple.
I do not cut out the extra candlesticks, as I do not add the missing))
In fact, instead, the time series shrinks or expands.
There is another interesting question. But what if we apply it not to daily candlesticks, but to five minutes.
Despite the fact that I consider the time series to be scale invariant, I also believe that for each scale it is necessary to build your own base with a pattern. And in the future I will go this way.
Then your and my question will not arise.
I conducted a huge number of experiments to recognize a wide variety of photos and screenshots. There have already been several implementations. At the moment, I settled on the following approach. I assume that by taking a screenshot the user can accurately and accurately cut out the area of interest. Conversely, a photo may have more errors, so it is subjected to more processing.
In general, I remove vertical and horizontal borders with unnecessary information. Random horizontal lines or indicators are also removed. But it’s better not to allow this.
I assume that a user who wants to get a quality forecast can train to make high-quality photos or screenshots in 5 minutes. It would be a desire.
Great idea!
I thought about it.
Thank!
This feature is usually built into the functionality of the smartphone’s camera.
In my experience, it is usually enough to switch between 16x9 and 4x3 mode, crop is not needed.
About this image. If this were a photo, then all unnecessary information will be deleted, including a horizontal line. If this is a screenshot, then I assume that the user left only the prices on the screenshot, although some of the extra information will still be deleted, but something will remain.
Recommendation for this image and others: Avoid hollow candlesticks. They should be painted in a color different from the background color.
This is not a signal system. These are statistics based on similar situations in the past. I did not train the system in the sense of learning. There is mathematics, there is a strict selection criterion. If you understand the issue of finding similarity in time series, you will understand what I mean.
The only thing I have to take care of is that the historical base be large. Now it is 170 thousand patterns. I think there should be several million. How to get such a base? There are several ways. But this is a completely different conversation ))
Thanks again for testing, feedback and questions.
All the best! Success in trading!
I will be glad if you decide to continue communication on this topic.
Yes, I will reply to your post later today/tonight. Thanks!
I believe a choice would be good. For example, allowing the user to pick the forecast timeline, from one to 21 (~30 calendar days) or more days.
Some users may trade retirement accounts that only allow so many trades per month, for example.
Rather than an average of the days, if allowed to choose the lookahead period, users might prefer knowing the stats for that exact lookahead day.
Not sure I understand. If the AI was trained on a certain number of lookback days, 40, for example, wouldn't it be best to give it the same exact number of lookback days in the photo in order to generate the forecast?
The last 40 days of an instrument will not look the same as the last 80 days "shrunk" down to fit into the "space" used by 40 days of the "training" data, no?
True, but users may be reluctant to give your app access to the 'all media on device' permission needed to use the screenshot (gallery) functionality, due to privacy concerns.
Even so, I suspect users with still use it for signals ... at least try to.
You're welcome. Keep up the good work!
%%
I still like a 200dmna[edit later with discretion];
+ grids which are simply a time saver more than an indicator/perhaps.
50 dma =ok;
WSJ use of 65dma is too little/too late , but it may help them sell papers...…………………..