Trial and Epic Fail

Ventures of an ex indie game developer

A final DLNN attempt

I should have done this ages ago, but finally I'm going to get around to it (as long as our new company business allows for it, I'm officially a used car salesman). DLNN, or AI as we say in the common tongue.

Here's what I'm thinking. I'd like to train the network similar to how Alpha Zero did their stuff. For features I'd like to use BTC and perhaps some other currency and also Bollinger Band % over a couple of thousand candles and/or SMA over some tens of thousands of candles. Then perhaps use a time frame of a couple of hundred candles or so as a training window. The additional instruments with a long time frame should contain information that is neither present nor deducible in the training window. Or so goes my thinking at least.

The network output could be buy or sell and possible neither. Or maybe how much. I'll use pytorch, which I haven't really used before (I've only ever understood and used tensorflow a bit). I'll also focus on leveraged trading, as that should be most profitable if the neural net can predict with a whiff of accurately.

The Alpha Zero has a game distinction which is harder in trading: either you win or you lose a game. It's never as clear-cut in the real world, but perhaps I could do something like take the lowest and highest "score" of a bunch of training windows and softmax the thing. Possibly.

Important is also to pick a time resolution which is high enough to avoid bumping into over-fitting. Or keeping the network small enough. One thing that is a lot easier than Alpha Zero is that I can do away with the Monte Carlo tree search.

Some of the hard things for me will be:

  1. Sizing and setting up the pytorch network;
  2. Saving+loading models;
  3. Trying out different input features;
  4. Transforming the output to trading activities;
  5. Quickly evaluating the profit of a training batch;
  6. Transforming profit into an error — and find a way to punish inactivity;
  7. Determining what is good evaluation data (do I even need it?);
  8. Trying out different constants such as maximum margin, or even concluding that those things might be better off as network output variables;
  9. Keeping track of everything, such as what output belongs to what time frame, etc.
  10. Speeding up training.
  11. Knowing when to stop training and ensuring the model doesn't over-fit.

A simple off-by-one error anywhere will be really bad. But if everything is done properly it should be well equipped to handle pumps, dumps and trading ranges, which is super difficult with plain math. I also think it should train and converge quickly, as trend-following works great, and during trading ranges it should hopefully transition to scalping tactics.

Anyhootch, a sink-hole of ignorance, bugs and optimization. Simply trial and error. Will let you know how it goes.

Crypto expectations

My last guess was right, crypto was on the rise and has now been for four months. From hereon I think price will go a little higher, say 66k or so in the next few days. Then there might be a month or two of reconsilitation and at the end a drop of 20-30%. After that, we're heading higher. Possibly in the 150k range.


That will probably also be the last leg where one can make silly money by following the trend. In four-five years time, BTC's price will resemble that of most other world currencies and only have smaller movements.

Crypto is back?

Last month or two, simple trading strats (including my bot) shouldn't have been losing any money. And perhaps the turning point from this stale mate is behind us. We'll need some more confirmation, but it looks like around October 15th, 2023 was the turning point.


Finally! ATM it looks like BTCUSDT futures is the best bet, but usually after a BTC pump we get the shitcoins disipation. S&P500 is down in the last couple of months, but US 10 year bonds yield just broke out, so who knows if this is a fake-out, but either way it probably is going to be a good run either on the bullish or the bearish side.

Driverless Teslas and remote viewing

Tesla's decision to not use any other means than cameras to form a view of the world were predicated on the knowledge that we humans only use our our eyes to do the same.

That view is called materialism, or physicalism, and has been lengthily and completely refuted by science. Btw, that's why the science elite are all into "woo-woo" in times when it's ok to admit it.

Anyway, a small portion of our daily life is probably composed by remote viewing — or better yet — sensing. I'm even guessing a small portion of each moment is composed of remote sensing, just as our other senses each help make up a whole.

I thought Tesla and others would be running self-running cars by now. But it seems like those that use lidar is actually doing better atm., which surprised Musk (and me). And sure, it's easy to make bad predictions in 2015 when 95% is done, and only 5% of the ice berg is remaining. But what if that wasn't it?


Even if a Tesla is able to circumvent all sorts of roadwork in the future, perhaps it will still be worse than humans in some situations. Not because of lack of data. Unique situations happen all the time, and a human who has never experienced it before is usually able to handle it fine. So data is not the issue. Perhaps remote sensing is? Remote sensing the future would be a great way to traverse most obstacles.

Fighter pilots' brains are wired differently than the rest of us. Either this is the physical reason they became fighter pilots, or that wiring helps with remote viewing, which in turn make them better at responding to novel situations in a timely and accurately fashion.

If this is the case for both car drivers and fighter pilots, it will take a long time to replace them with a data-driven design. How long exactly? I'm guessing 10 years.

ChatGPT-3

It keeps blowing me away. Haven't gotten access to GPT-4 yet. Bah! But just look at this stuff:




 Thinking of making a couple of easy language model games. Hmm.

Free bot trading required

See if you can spot the time when Binance turned on fees on BTC/USDT, and turned them off on BTC/TUSD.


Trading volume instantly up by a factor of 100. My bot "Arrow" had been doing so well for the 11 days it was allowed to run: 4 - 55% per day! No more such good stuff, unforch.

Will have to make do with the crappy BTCTUSD instead. In a few weeks I expect volume to go way up, as more and more HFT bots crave free trades. When it's high enough, they'll turn fees on again. Murmur...

Incontinetia Buttocks

Just because I'm paranoid doesn't mean they're not out to get me. As usual, here's backtesting of my latest and greatest trading strat:


Note that soon after I launched comes the first "recession." All the other downhill slopes are harsh, but this last one is more gradually inching down — although there's not much price action.

This is not the first time this has happened.

The question is if my backtesting is overly precise. Like overfitting for an ANN. Or if my paranoia isn't paranoia at all.

My one minute trader is doing a lot better now though, since it has five years worth of high quality data. But lately it's just been holding steady (if you think +90%/-35% is steady, that is).

Whereas this one second trader has a much shorter period, albeit a lot more data. Six months price data is 15,552,000 candles.

So the question is if I scrap this idea, or simply put it on hold until the time is right?

Hard to tell. When things are going well, I feel like Biggus. But now I feel more like Incontinentia. Since the exchange supply is lower than ever (both BTC and ETH), I think my Biggus times are coming back soon.

About the author

Mitt foto
Gothenburg, Sweden