Machine learning in trading: theory, models, practice and algo-trading - page 107
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
Ternary - means that it can take three mutually exclusive states. Another name is ternary.
And a grid with three outputs, each of which is binary, can produce 8 mutually exclusive states, of which only three are interpreted unambiguously, like ternary. And it is not clear how to interpret the remaining 5 states?
I also use ternary in some way, I have three classes - up-turn, down-turn and no-turn, those are 1,-1,0
i.e. in this case it is not necessary to make a quality forecast, you may earn not by making a good one but by using a small stop and large profit.
it's not a system, it's just a generator of entries with stops
but the sad thing about this approach is that it's not clear how to train the model
===========================================
Despite the fact that the model works weakly, it does not prevent it from making stable profits, and 14% a month is not the limit, I've seen 35%, it all depends on how to train the model
=====================================
And the earnings at the expense of the stop/stack ratio of 10:1 is quite stable, plus there is risk control
I also use ternary in some way, I have three classes - up-turn, down-turn and no-turn, those are 1,-1,0
And in principle with this approach it is not necessary to make high-quality forecasts, you can earn not by making high-quality forecasts but by taking small stops and large profits, i.e. due to risk management
it's not a system, it's just a generator of entries with stops
but the sad thing about this approach is that it's not clear how to train the model
===========================================
Despite the fact that the model works weakly, it does not prevent it from making stable profits, and 14% a month is not the limit, I've seen 35%, it all depends on how to train the model
=====================================
i got a good feedback! i got a good feedback when i looked at the chart and i got a good feedback when i looked at the chart. i got a good feedback when i looked at the chart. i got a good feedback when i looked at the chart.
Well, that's great! What's the problem?
I do not understand how to teach such a model, all methods are sharpened to the error classification and in my approach, the probability of guessing will always be below the plinth, in short, the effectiveness of my model must be assessed differently, but how do not know
I do not understand how to train such a model, all methods are sharpened to the error of classification, and in my approach the probability of guessing will always be below the plinth, koroch the effectiveness of my model must be assessed differently, and how do not know
So you're saying that 14 percent a month, and sometimes you get as much as 35. Then don't worry any more about the obscure, because the results are amazing, without false modesty for the results.
Although, I wouldn't use fixed stops to exit the market (as we know patterns have a tendency to scale any way before and after entering the market). But for simplicity I sometimes did so: I trained for an exit with sl/tp=1/3 stops, but used a 1/2 ratio on OOS (gave a head start to the neural network, for which it was grateful in the form of increased number of correct responses compared to what it would have been if I had used 1/3). Como, as I said it is necessary to limit the deal in time, because the probability that in the future the price will not reach SL and TP ever, though small, but still there, and you can not say that the grid trained bad, can only say that life is short.
Somewhere there was an idea that the model should predict the movement of at least a specified number of points during training... This is a very sensible idea in my opinion...
I can add - a movement of at least a specified number of points taking into account the typical volatility at this moment of time in the appropriate timeframe and the specified time of trade life. The last time I worked on grids two years ago, I was working in this direction. In general, I abandoned grids, because some probabilistic characteristics of the market were not clear to me. Now everything has settled more or less in my head in this matter and probably it is worth to continue working on the grids....
As it seems to me, the role of "machine learning" methods in TS should be reduced to the minimum possible, and the market factors that stably repeat from year to year should be brought to the forefront. For example, does anyone in this thread use the knowledge that the volatility is maximal in the middle of the trading day? - unlikely.... And this is an undeniable characteristic of the market, which does not change.
There is another observation (invariable fact) known to all, but stubbornly ignored by "machine operators" - on lower TF the price behavior during the night time is sharply different from the day time.
But this difference comes to naught on TFs larger than H1, and isn't that why many TS show more stable results on higher TFs (because the candlestick price change is more or less homogeneous)?
But we want more deals (with inevitable bigger losses due to commissions and spread), that is why we have to use TFs lower than H1. There are two ways of solving the problem of "different price behavior" within twenty-four hours: 1) or TS division into "night" and "day", 2). Or limiting the time of day trading (for example from 05:00 to 20:00). Usually I do not take this trouble and go by the second variant, but even such a simple filter by time significantly improves the results of learning and further trading.
For "night" intraday I haven't managed to build an adequate TS on neurons, because there are other, imho, rules different from pattern combinations.... Which ones - is a big question, but the main question is whether the grids should be applied in night hours (rules for which, predictors, are difficult to formalize), if in these very night hours can be successfully applied a simple and uncomplicated TS, such as trading in a channel and similar variations on the theme of channels ...
There is one more observation (invariable fact) known to all, but stubbornly ignored by "machine operators" - at lower TF the price behavior at night time is sharply different from the day time.
Time should be included as predictors (along with others).
Each of predictors is divided by the corresponding number of predictors, e.g. hour number by 24 predictors. For example, the first predictor has 1 for the 1st hour and zeros for the other positions. The second predictor has 1 for the 2nd hour and zeros for the other positions, etc.
If we check the predictive ability of such predictors, it turns out that each such artificial predictor has a different predictive ability. For example for a day of the week it is Wednesday and Thursday. The other days of the week = noise and should be excluded from the model.
That gives very good predictors.
Time should be included as predictors (along with others).
Each of the predictors is divided by the corresponding number of predictors, e.g. hour number by 24 predictors. For example, the first predictor has 1 for the 1st hour and zeros for the other positions. The second predictor has 1 for the 2nd hour and zeros for the other positions, etc.
If we check the predictive ability of such predictors, it turns out that each such artificial predictor has a different predictive ability. For example for a day of the week it is Wednesday and Thursday. The other days of the week = noise and should be excluded from the model.
We get very high quality predictors.
Exactly. Almost. The hour number just needs to be broken down into 23 binary variables...
And for some methods, you don't need to do that either. A random forest will handle that cat variable by itself.
There is another observation (invariable fact) known to all, but stubbornly ignored by "machine operators" - on the lower TF price behavior at night sharply differs from the day.