Bitcoin Forum

Hi everyone,
we just launch the beta of Presage a deep learning system use to forecast cryptocurrency market price.

You can learn more about the product at: https://www.presage.tech

Example result from the beta:
https://i.ibb.co/TbtM9JM/beta.png

We are collecting all the feedback you may have on the wish list or contact us via the chat.

Have a great day !

I quickly checked and it seems it's more like an HI doing the tasks instead of an AI since adding a pair, always adds 180 minutes to its delivery time...
^{- If it's an AI, in both theory and practice, it should be able to perform multiple side-by-side tasks with little to no added time in regards to its delivery time.}

---

I noticed there's no discount when adding more pairs ^{[it's always +$70]}...
^{- You might want to reconsider that part and also offer a trial of some sort.}

I will try to not give too much about our architecture. The system rebuild all the weight from scratch for every client. There is not multiple side by side interaction, we are trying to rebuild the rule of a non turing complete automaton. We are limited to 16 order everyday using a Tesla v100, each order need 3h of computation. (per crypto pair)
If you have any question feel free to ask. Don't make any assumption on something really complex.

We will be able to reduce significantly the price when we reach this first step of max capacity.

Have a great day!

Ps: thank you for your interest :)

Your website claims to have 94% accuracy. It doesn't specify the metric or methodology that caused you to come to this conclusion, but this sounds phenomenal. If this could be translated into auc (https://developers.google.com/machine-learning/crash-course/classification/roc-and-auc) of 0.94 for investment choices, your model would be very valuable.

With that being said, I think it is more likely you are either using flawed methodology, or some of your training data is leaking over to your test data, causing overfitting on your test data. Your testing procedure may also be otherwise flawed. I would encourage you to review your training and testing procedure, and re-evaluate your claimed 94% success rate.

The metric to calcul the accuracy is directly done in your browser. The javascript is not obfuscate. We are comparing the distance between forecast value and market value.(the market value are update every hour)

Now, the knowledge you are referring to, apply for classification. Not for forecast. Overfitting and underfitting are not a concept you can use to predict future value. I invite you to try to forecast future value to understand the phenomenon.

I invite you to try the product and not speculate on it.

Ps: the model can spike to 99% accuracy but some dip are not predictable

Thank you for your interest :)