Artificial intelligence predicts which levels will be broken or held based on raw market movement, EMA, HMA, MACD, Bollinger Bands, Keltner Bands, ATR, RSI, and Ichimoku cloud. The success rate is 73%.

The thick line labels the level to be broken or held. The rectangle below or above the thick line defines the price range where the prediction is valid. Note, there are 3 neural networks for time frames 1h, 4h, and 1d.

How to trade? Enter at a level in the rectangle where you expect that the price bounces away to the direction which is predicted by the AI. Exit when the price leaves the rectangle.

Success rate? The following table shows success rates on new data for particular probability or higher of neural network prediction. This means, for example, that if we select probabilities that are higher than 95%, we get a success rate of 78.64%.

Probability cond. | Success rate | All samples | Correct | Incorrect
> 95% | 78.64% | 75606 | 4076 | 1107
> 90% | 77.46% | 75606 | 7594 | 2210
> 85% | 76.79% | 75606 | 10774 | 3257
> 80% | 76.45% | 75606 | 13865 | 4272
> 75% | 76.21% | 75606 | 16774 | 5236
> 70% | 76.12% | 75606 | 19687 | 6175
> 65% | 75.98% | 75606 | 22515 | 7119
> 60% | 75.94% | 75606 | 25330 | 8024
> 55% | 75.84% | 75606 | 27935 | 8898
> 50% | 75.74% | 75606 | 30687 | 9828
> 45% | 75.63% | 75606 | 33364 | 10749
> 40% | 75.46% | 75606 | 35959 | 11694
> 35% | 75.23% | 75606 | 38448 | 12656
> 30% | 74.98% | 75606 | 40978 | 13676
> 25% | 74.76% | 75606 | 43442 | 14667
> 20% | 74.46% | 75606 | 45888 | 15738
> 15% | 74.28% | 75606 | 48402 | 16760
> 10% | 73.96% | 75606 | 50961 | 17945
> 5%  | 73.71% | 75606 | 53424 | 19056

Training dataset?

We trained the neural networks on 75% of our dataset (read more about our dataset below), which is around 655 years, and we tested it on the other 25% (219 years)—this means that they achieved a success rate of 73% on new data.

We test our algorithms on 67 Forex assets from the year 2000 or later depending on availability and Bitcoin from 2011.

It is 874 years of data in total (from 1 minute TF). Testing on such a huge dataset requires a lot of hardware—we have a dedicated server with powerful CPUs and GPUs installed just for the purpose of backtesting, forwardtesting, and training.

Applicability?

Note also, the AI is trained mostly on the Forex data, which means that the success rate may differ if the AI is used on the other asset types (e.g., cryptocurrencies)