Ideas for managing fees

[owocki]

• Integrate with a low-fee or no-fee exchange.
• Raise the threshold for which the trader decides to buy/sell, and lower the threshold for which it decides to hold.
• Manage fees when pytrader takes multiple positions in the same instrument in different directions

[NDuma]

Taking advantage of Maker/Taker fee structure will produce an edge. BitFinex's Maker fee is as low as 0% while Taker fee is on par w/ Poloniex. This requires reading the Order Book to find gaps & similarly / the same Optimal Placement w/in 'Depth of Market'

[owocki]

This requires reading the Order Book to find gaps & similarly / the same Optimal Placement w/in 'Depth of Market'

Also, a good way to track the open order and decide whether / how long to wait for it to be executed.

[joeljuca]

FYI: BTCC doesn't charge trading fees. Supporting it can be interesting - and profitable.

[jeff-hykin]

Is there any reason poloniex was chosen first?

I was talking with one of my friends (a young investor) and we realized a 10% reduction in fees would’ve gotten statistically significant profits (would’ve ended with 1.2 BTC).

Without fees it would’ve had a doubling rate of less than 1 month, in which case 1 year of trading would be an extrapolated 4,100 BTC ($1.7 Million).

So I’d say supporting no-fee trader sounds like the best way to reach profitability.

[hughstephens]

BTCC looks to support CNY and BTC (Litecoin too?) only – not ETH? Unless I'm missing something? Working from https://exchange.btcc.com/ – they also have a pro exchange but looks to only support CNY/BTC pair. They do support websockets stream of market data though which is a plus (https://www.btcc.com/apidocs/spot-exchange-market-data-websocket-api)

Agree that fees are going to be a massive element to this, but I'm guessing that could be why Poloniex was chosen first – supports the chosen pairs. Maybe we need to put together a doc of all the trading platforms that have an API, support XYZ (atm ETH, BTC, USD?) and their fees?

[jeff-hykin]

Here’s the details for BTCC ( from BTCC itself )

BTCC also seems to have good API support

I think a comparison document would be helpful

[owocki]

Is there any reason poloniex was chosen first?

Not a very defensible one -- It's where I trade my BTC/ETH manually, and they had a python API client.

I agree that fee management will be a major part of getting to profitability. Assuming that this repo can get to a point of competing for profits before folks with more computing power / quant knowledge enter the market.

[owocki]

I was talking with one of my friends (a young investor) and we realized a 10% reduction in fees would’ve gotten statistically significant profits (would’ve ended with 1.2 BTC).

Without fees it would’ve had a doubling rate of less than 1 month, in which case 1 year of trading would be an extrapolated 4,100 BTC ($1.7 Million).

Was thinking about this on the bike ride home tonight @jeff-hykin. Something I need to make clear to you, in the spirit of transparency.

(a) While building this repo, I went through a couple trading methods:

1. Buy (and accumulate) a coin over time until the ML libs predict a downtrend, then dump it over time.
2. Buy a coin and then immediately trade it back after granularity minutes.

For much of the time this repo was live, I was holding a significant amount of ETH.

(b) I built this repository during one of the most phenomenal price rises in cryptocurrency:

TLDR -- I have not done an thorough analysis of whether I was lucky or good at trading during this time. When you project sums like $1.7 million, you are getting ahead of yourself.

We should focus on proving repeatable results before we get dollar signs in our eyes. Is ETHs rise repeatable? Probably not. Is ML a viable tool for predicting crypto price trends? Probably. Is there work ahead of us to get there? Probably.

[jeff-hykin]

Yes I agree despite 23,000 trades and 2 months, there’s still a large void of proof. And yes $1.7 million is unlikely to get even with a weekly doubling rate (from any source) due to inherent unsustainablility. None the less it would’ve been an impressive growth rate.

Although in hindsight it seems somewhat obvious, I didn’t realize ETH was a major coin being held by PyTrader. That does change the meaning of the results a good bit. I was curious how the buying tactic worked, I would’ve assumed more of the 2nd tactic (buy and then sell shortly after).

I do think an evaluation of actual profitability would be valuable. Although the “Managing Fees” section feels like wrong place to be posting this, I’ll at least mention it.

A quick way would be create a theoretical control group: (if you bought 1 BTC/ETH and held it for the same two months) and compare the profits.

A second much more rigorous method (what I would do before even investing 1 real BTC) would be to get a collection of data from different markets and simulate PyTrader at random segments on them. This would be a true experiment and give a good idea of the actual profitability of the algorithm in the general market. But, I imagine this would take quite a bit of computation power. Do you have a general estimate of how long it would take a $900 PC running 24/7 to simulate 2 months of trading? (assuming all the code is modified somehow for this to happen)

There should probably be a well documented evaluation section before any specific profitability projections are discussed with any seriousness.

[owocki]

I would’ve assumed more of the 2nd tactic (buy and then sell shortly after).

I can see why you'd assume that -- The repo, when I open sourced it, exclusively does this (see trade.opposite_trade on the trade model). I added this logic sometime in late February / early March 2016.

a collection of data from different markets and simulate PyTrader at random segments on them

Part of the reason I didnt do this is that one would have to simulate the exchange and the filling (or partial filling, or not filling at all) of orders. I am not quite sure exactly how to simulate the logic of whether an exchange fills an order, but am open to suggestions.

Do you have a general estimate of how long it would take a $900 PC running 24/7 to simulate 2 months of trading?

It depends on how frequently the trade is making trade decisions. If the analysis is every 5 minutes, we'd need 12x the compute power than if it's every 1 hour. If I had to put my finger in the air and make a guesstimate, I'd have to think we could simulate 2 months of trading for 1 hour granularity on the order of minutes, maybe hours, with the present codebase.

There should probably be a well documented evaluation section before any specific profitability projections are discussed with any seriousness.

Agree. I've begun tracking a 'further analysis of initial results' issue @ https://github.com/owocki/pytrader/issues/32 and an issue for evaluation of trading models @ https://github.com/owocki/pytrader/issues/33

[owocki]

I will have PR ready after work today in which I begin attempting to manage the fees. I'll post it around 6pm MST if anyone on the thread would like to code review.

[owocki]

https://github.com/owocki/pytrader/pull/60

[rubik]

Here’s the details for BTCC ( from BTCC itself )

BTCC looks promising, but I believe that you can only trade BTC, LTC and CNY. Also, for leveraged trading the fees are different: