AI talking | Page 6 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

which fibonn was vegas talking about? 1 reply

Changing Your Self Talk By Talking to Yourself 2 replies

My left brain starts talking to me 7 replies

talking to myself(smile) 8 replies

Attachments: AI talking
Exit Attachments

AI talking

  • #101
  • Feb 14, 2023 6:57am Feb 14, 2023 6:57am
Quoting joyny
Exciting results.. turns out self learning AI is possible and the longer AI trades/learns the better results it provide. how I tested this: 1. Made 45000 H1 candles training data 2. Split data by 100 H1 candles so AI trained first iterations with 1000 H1 candles, then forwardtest H1 100 candles (~4 trading days) 3. After every 100 H1 candles retrain AI model - add new candles to training data so training data every iteration enlarges by 100 Here full AI self-lerning iterations audit: {file} Here some insights.. In the beginning, we can notice bad...
so simply you train with history data and then test on more history data where its new foe AI, right?
WaBaS Where are Buyers and Sellers
  • #102
  • Feb 14, 2023 6:58am Feb 14, 2023 6:58am
Quoting joyny
seems to need to play around with those too... a lot of analysis needs to be done here.
made first comparision DecisionTreeRegressor vs RandomForestRegressor

Attached Image (click to enlarge)
Click to Enlarge Name: DecisionTreeRegressor_forwardtests.png Size: 32 KB

Attached Image (click to enlarge)
Click to Enlarge Name: RandomForestRegressor_forwardtest.png Size: 32 KB

huge difference in correct predictions %.. and 2x fewer deals which is promising.. will prepare additional demo account and run in parallel those 2 AI models..
  • #103
  • Feb 14, 2023 7:07am Feb 14, 2023 7:07am
Quoting joyny
What we see from AI predictions: 1) GOLD - UP - according to AI OK 2) GBPUSD - was a gap on Monday opening up and now GBPUSD indeed is going UP OK 3) Bitcoin - you all see what's happening... UP too. OK 3 audited AI predictions all 100% right for this week so far. Nice job my AI "mentor"...
direction prediction is not big issue, its the DD%
WaBaS Where are Buyers and Sellers
  • #104
  • Feb 15, 2023 3:07am Feb 15, 2023 3:07am
First AI version traded on Demo account allready 38 hours. Did 17 trades. 8 profitable. Was losing serie 5 in a row. Win serie 7 in a row.

Noticed that Python script execution usually takes 3-4 seconds. Was once 18 seconds and 9 seconds. Maybe on a better computer that could work more stable.. but ok this did not impact trading results so far. The first deal was with a mistake in EA SL/TP calculation.

Attached Image (click to enlarge)
Click to Enlarge Name: 38hours.png Size: 50 KB

Compared prompt messages with those exported in CSV - equal. This was my biggest concern that I did some mistakes with timings shifting while exporting CSV and when running a simulation of signals from CSV file. But seems now this is correct.

As AI trades patterns then it is essential to ensure that prompts in training are the same structure (creation rules) as in live trading. For a pattern is important if used the current candle or shifted 1 previous candle. As I see this is equal and EA prepares correct prompts for AI predictions.
  • #105
  • Feb 15, 2023 4:26am Feb 15, 2023 4:26am
Quoting joyny
First AI version traded on Demo account allready 38 hours. Did 17 trades. 8 profitable. Was losing serie 5 in a row. Win serie 7 in a row. Noticed that Python script execution usually takes 3-4 seconds. Was once 18 seconds and 9 seconds. Maybe on a better computer that could work more stable.. but ok this did not impact trading results so far. The first deal was with a mistake in EA SL/TP calculation. {image} Compared prompt messages with those exported in CSV - equal. This was my biggest concern that I did some mistakes with timings shifting while...
if the loss is due to costs (comission and/or spread) you should optimize, if not then test reverse trading with no change
WaBaS Where are Buyers and Sellers
  • #106
  • Feb 15, 2023 4:38am Feb 15, 2023 4:38am
I am 35k+ hours screen time experienced trader/coder. if you want to collaborate tell me
WaBaS Where are Buyers and Sellers
  • #107
  • Feb 15, 2023 4:49am Feb 15, 2023 4:49am
Here AI2 (with Random Forest Regression algorithm) forward tests results (prepared signals CSV with Python and then executed in MT5 with real ticks):

Attached Image (click to enlarge)
Click to Enlarge Name: ai2_forwardpng.png Size: 20 KB


Attached File(s)
File Type: xlsx forward_results_v3.xlsx   1.8 MB | 246 downloads

It is what it is... cant comment.. I know I know..

Demo will show.

With this AI model, the biggest risk could be the time of Python file execution. Because it took much more time for training.
  • #108
  • Edited 5:31am Feb 15, 2023 5:16am | Edited 5:31am
Quoting joyny
Here AI2 (with Random Forest Regression algorithm) forward tests results (prepared signals CSV with Python and then executed in MT5 with real ticks): {image} {file} It is what it is... cant comment.. I know I know.. Demo will show. With this AI model, the biggest risk could be the time of Python file execution. Because it took much more time for training.
Time of Python execution may be a non-issue because on the timescale of seconds it probably is a wash over time. 50% of the time it increases a loss a bit, 50% of the time it will give you a little more profit. It should be a random wobble and just have zero net effect over the long-term.
So your demo results show about 50% win rate right now. You need more real world candles to know for sure. For now, you could look at your backtest data, take 38 hour chunks within it and see what your average win rate variation is for 38 hour chunks historically for comparison.
  • #109
  • Feb 16, 2023 10:17pm Feb 16, 2023 10:17pm
Quoting joyny
tests results.
How's the demo trading going so far? You should have data from about three days now?
  • #110
  • Feb 17, 2023 12:29am Feb 17, 2023 12:29am
Quoting costos
How's the demo trading going so far? You should have data from about three days now?
AI1 results so far:

Attached Image (click to enlarge)
Click to Enlarge Name: IMG_20230217_082641.jpg Size: 221 KB

Was 8 losing serie.. max in tests was 6.. not good. Total dd is still acording to tests.

AI2 detected errors fixed yesterday therefore no enough correct data. For 1 day 66% profitable..
  • #111
  • Feb 17, 2023 4:46am Feb 17, 2023 4:46am
Quoting joyny
Was 8 losing serie.. max in tests was 6.. not good. Total dd is still acording to tests.
Here are historical stats for AI1 regarding drawdowns time:

Attached Image (click to enlarge)
Click to Enlarge Name: ai1_dd_time.png Size: 36 KB

As we can see it is ok to be in DD for 1 entire week.
  • #112
  • Edited 5:34am Feb 17, 2023 5:24am | Edited 5:34am
Quoting costos
For now, you could look at your backtest data, take 38 hour chunks within it and see what your average win rate variation is for 38 hour chunks historically for comparison.
Did that.. for days till 16 February included it produced various results not seen at demo account.

Each time AI is executed it produces different signals.
Attached File(s)
File Type: xlsx eurusd_ai1_test_1week_try1.xlsx   38 KB | 217 downloads
Attached File(s)
File Type: xlsx eurusd_ai1_test_1week_try2.xlsx   36 KB | 221 downloads
Attached File(s)
File Type: xlsx eurusd_ai1_test_1week_try3.xlsx   37 KB | 222 downloads
Attached File(s)
File Type: xlsx eurusd_ai1_test_1week_try4.xlsx   39 KB | 273 downloads

When try enough times I bet will get some loosing too..

AI1 I wont use. Still keep it run on demo but it is not a good algorithm for this purpose. It returns 1 or 0 but not 0.74 or -0.43 as it do AI2. And in case of identical prompts it will return random result.

Will compare this with AI2 after 1-2 weeks.
  • #113
  • Feb 17, 2023 6:37am Feb 17, 2023 6:37am
Quoting joyny
{quote} Did that.. for days till 16 February included it produced various results not seen at demo account. Each time AI is executed it produces different signals.{file}{file}{file}{file} When try enough times I bet will get some loosing too.. AI1 I wont use. Still keep it run on demo but it is not a good algorithm for this purpose. It returns 1 or 0 but not 0.74 or -0.43 as it do AI2. And in case of identical prompts it will return random result. Will compare this with AI2 after 1-2 weeks.
Identical prompts resulting in different results is normal. If AI is performant it will perform on average well, even if individual runs will be slightly different.
These 4 test runs look good - kind of the same as what you saw in demo backtesting. Is that on forward test on real data?
  • #114
  • Feb 17, 2023 8:28am Feb 17, 2023 8:28am
Quoting costos
Is that on forward test on real data?
Current week. Compared with demo trades made this week and noticed that demo now in dd -8usd but all those 4 trys on same week ended up in profits.

1 week is too small timespan to conclude something.
  • #115
  • Feb 17, 2023 8:44am Feb 17, 2023 8:44am
Quoting joyny
{quote} Current week. Compared with demo trades made this week and noticed that demo now in dd -8usd but all those 4 trys on same week ended up in profits. 1 week is too small timespan to conclude something.
Irrelevant they are all different. If they all start on the same millisecond, all execute exactly together (if forward testing), and all use the same random seed in the same model they will be the same. Not what happens, right?
  • #116
  • Edited 9:18am Feb 17, 2023 8:58am | Edited 9:18am
Quoting costos
Not what happens, right?
Yep...

Quote
If the variance is large and results in a significant difference in the profitable trades percentage, then it may indicate that the model is overfitting to the training data or that the decision tree algorithm is not well-suited for the specific problem at hand.

Quote
To assess the variance in the results, you could calculate the standard deviation or coefficient of variation of the profitable trades percentage across the different runs with the same training data. Depending on the specific application and level of precision required, an acceptable range of variance may be between 1-10%, but again, this would depend on the specific context of the problem.
  • #117
  • Feb 22, 2023 4:49pm Feb 22, 2023 4:49pm
Quoting joyny
{quote} AI1 results so far: {image} Was 8 losing serie.. max in tests was 6.. not good. Total dd is still acording to tests. AI2 detected errors fixed yesterday therefore no enough correct data. For 1 day 66% profitable..
Can you share the myfxbook link?
  • #118
  • Feb 23, 2023 1:24am Feb 23, 2023 1:24am
Finally found the error - it was CSV file export for training data not correct. It included the "current candle" already in prompt data therefore AI was so easy to "learn" to spot that pre-made entry. Fixed.

And now all look much more realistic...

1. AI models scores are now negative (what is "bad")

2. No more yummy entries with a probability 0.7 to 0.9. Max probabilities now 0.5 and very rare. Normal probabilities are now 0.3 to 0.4

Attached Image (click to enlarge)
Click to Enlarge Name: fixe_ai.png Size: 22 KB

Here forward testing GBPUSD with unknown data for AI from 2019 january. Now it looks realistic. Still, 66% profitable deals with r/r 1/1 is good.

Anyway a lot more analysis needed to be done.
  • #119
  • Edited 4:17am Feb 23, 2023 3:56am | Edited 4:17am
Quoting joyny
And now all look much more realistic...
Here how AI drawdowns look.. 80 days sit in DD.. that's real trading

Attached Image (click to enlarge)
Click to Enlarge Name: dd_days.png Size: 9 KB

with 100 usd capital DD 30% - real life too.

here example of one of my live accounts trading >3 years now, was 2x sitting in ~6 months DD. That's real life when trading patterns which profitability % fluctuates over time..

Attached Image (click to enlarge)
Click to Enlarge Name: dd_examples.png Size: 15 KB

Cool. AI shows consistent results 4 years in a row. In case if I made no more mistakes.. trained with H1 data from 1996 till 2019. Then retrained after 10000 candles (about 80 weeks). Should run tests with more frequent re-training (time-consuming).
  • #120
  • Feb 27, 2023 2:35pm Feb 27, 2023 2:35pm
Won't use python AIs.. no stable results found.

I will use the simple neurons which is possible to implement directly in the MQL5 expert advisor and use optimizations to find neuron weights.
0 traders viewing now