wtf o_0

anyhow - algo doing bit badly. an ugly, prolonged drawdown

tweaked a bit, slightly lowered the bet

lo and behold - half way there to breaking even with a single trade



bugcoin doing funny moves

this is beautiful

I love Wednesdays



should finally do something about those pesky weekends

sort of bummer knowing they almost always red

------

1-2 weeks
scale up if breaks even / in profits

------

backtests makes zero sense



what the hell is going on?

yet to see a good use for windowing

I guess my approach just isn't compatible

What's the deal with these insane backtests? 1000+% annual return, but it sounds like the live results don't live up to this, if I read that correctly?

I think that's the issue with cryptocurrencies. Their extreme volatility makes it pretty much impossible to create reliable backtests.

Well... That's the thing with backtests. Past doesn't repeat itself.

Green backtest is just a small step to trade effectively.
Markets aren't simple - there's a lot of stuff going on.

Trading fully automatically while basing decisions purely on TA (with no access to volume) is quite ambitious.

About 600% from those ~1200% are from May crypto crash which is a black swan and doesn't repeat monthly.
I do think it's important to include ability to surf such waves.

Out of those 600% - I would say 50% to 200% *might be* sort of actually realistic.

One issue with measuring performance is that I'm constantly tinkering with it.
However - I do believe in such approach. If you keep bashing squares into circled holes - eventually they become circles.
Backtest shows PF1.7 (on a flat bet) but I see PF0.9, maybe PF1.2

October was in slight profits. But there were like at least 10 iterations of algo updates (starting from about AR 300%).
And mostly on fixed $10 trades.
November is red so far.
Markets retesting all time highs, stuck in indecisiveness. Probably not for long.



Even if history did repeat - soon enough it wouldn't.
Pulling out 5% of total market cap can't go unnoticed.

But you never know unless you live trade.

Also - it's important to understand that 0% -> 10% is not the same as 1000% -> 1010% on a simulated compounding system.

Those numbers get ridiculous really fast.

I don't know what your strategies and workflow look like but "tinkering" here sounds a lot like overfitting, doesn't it? If you're adjusting parameters or adding new mechanisms that increase the AR in the backtest, you can introduce all sorts of biases into it even with OOS testing.



You have to be confident that your strategy wins this black swan event not by chance though. I've also played around with the MATIC coin some time and the backtests were often defined by that burst of volatility in may. But if there are 2-3 trades happening that capture these insane price jumps, I just don't know how I can know if that is random luck or if the script can actually be on the right side of these reliably. And since I'm pessimistic, I rather exclude those times from the backtest personally. Though I also don't include any mechanism in the scripts so far that specifically react to these moments.

Yes and no.

I mean - I would gladly have PF0 algo that somehow magically is vastly profitable.
Question is - how do you know?

How do you determine this mythical boundary at which improvements suddenly becomes overfitting?

5 successes out of 10 are much less meaningful than 500 out of 1000.
One of my rules is to strive for relatively high trade count (>2000).

I have removed couple weird rules despite losing %.
I still have some of them.

Trade and see. Adapt. Think deeply about every rule you adjust.
It's not random.

You can enter a long, see a dip and think - it's over.
Or you can see it as a temporary sweep, wait a bit longer and be in profits.





I am reasonably confident. It's not just 2-3 trades. There's like hundred in May.
Would say - it's actually algo's strong suit to handle such times.

Marked other bursts of volatility.
Either it rides it or gets stopped out and stays relatively flat.



Mentally - I do that.
Practically - I would rather have them just to keep some sort of assurance whole thing won't go bankrupt in seconds.

Keeping such ability seemingly hinders algo, removes a lot of room during "normal times".
But on grand scale of things - I think it actually rises the bar and increases adaptivity.



It's important to think about synergy.

I try to build it from "you don't want to flip here" standpoint instead of "this is the time you trade".

Instead of SMA 50-200 cross or perhaps valley/peak like rule - I favor rise & fall.
It's careful filtering instead of catching glaringly obvious tells (in fact - I might include such layer in future).

When you build it such way - rules seem to indicate same thing yet do that differently.
Once you get to combinations of rules - they start to cover up each others weaknesses.

If we take that same SMA 50-200 cross and combine it with (imo shitty) UO - there's a good chance you can change it to 49-199 cross.

Well, I think we have very different styles to be sure

Like, if your strategy actually works the way it appears in the backtest, then you will be filthy rich quickly. But if it doesn't, if you don't know to what extent it is biased/fitted, if you don't know to what extent it represents reality, I wouldn't use it at all to judge my strategy.