Share: Rebalancing Premium in Mean-reverting Cryptocurrencies

Awesome share thank you Emi

Emi Gal

Thank you for sharing.

It's good to know that selling winners and buying losers someone called a "rebalancing premium".

Other notes:

Daily rebalancing of 25 instruments over 2 years should create a lot more than 331 trades.

What if we rewrite line 100 like this?

                    if self.Portfolio[symbol].Invested:

Vladimir - thanks for the comments.

If the assets are mean-reverting, selling winners and buying losers should generate profits, as the assets will be reverting back to the mean. At least in theory. 

Great point regarding the number of trades. Seems this algo gets killed by fees when I make the change you suggested. 

Emi

Emi Gal

Here is a modification of your algorithm, mainly related to risk protection in  DailyRebalance(self).

Hope you will enjoy “rebalancing premium”.

Vladimir that’s a great improvement, thank you. 

This is an interesting idea. I re-ran this and it looks like after the recent general crypto market downturn, this algorithm still does a lot better than a buy-and-hold BTC strategy (468% vs roughly 317% for BTC).

A few things I noticed:
- during market downturns the algo seems to suffer significantly more drawdown compared to BTC (37% vs 35% from May 6-21 2021, 34% vs 23% from Dec 19 2021-Jan 26 2022, and most recently 54% vs 37% from March 28-May 11 2022)

- during longer uptrends in the market the algo seems to outperform BTC (322% vs 278% from Nov 2 2020-Feb 19 2021, 136% vs 29% from June 30-Sept 6 2021, and 49% vs 46% from Sept 20-Nov 8 2021)

I wonder if trying to find a risk-off condition based on long term trends would help to mitigate some of the drawdown?

probably add a simple MA parameter don't trade under 100 200 50 ect

karamazov - I was planning to run the algo on Q1-22 to see how it does under a significant market drawdown. Thanks for doing that, and glad to see it did well! 

In terms of drawdown - Elsid Aliaj is right, a SMA does the trick. I've actually built a version that lowers drawdown to ~34% by not trading under SMA. Ran an Optimizer on it and seems the 40-day SMA provides the best CAR / DD ratio. Will try to polish it up & publish this weekend. 

Indeed it does Elsid Aliaj . Performance improves significantly when using a simple 25-50 EMA crossover indicator. 

Also interesting is that the algo actually performs a little better and has slightly less drawdown without using the 30% stop loss condition.

Emi Gal maybe also test, selling winners only when below MA, only buy losers when above MA, and vice versa to see what sort of implications it has. Not buying and selling at the same time.

Also realistically 2.7 sharpe is probably unsustainable if you need to lower drawdown cut your account size by 50-75% and target what drawdown you want. 

Emi Gal

Here is another less expensive version. 

Hi everyone,

Attaching a new version that only trades if benchmark price (BTCUSD) is > 40-day SMA. Performance is lower than Vladimir's version, but with significantly lower drawdown (32% vs. 66%). 

Emi

I have tried replacing BTC with ETH, it produces higher results. Downsides - somewhat higher drawdown and way higher fees.

hi Emi,

I'm new to QC and have found your algorithm very insightful.

When replicating it I cam across two observations, which might or might not be down to my lack of experience with QC:

1. The algorithm seems to start rebalancing before price feed is available for all assets. I'm getting error messages along those lines:
    

Backtest Handled Error: FTTUSD 2MN: The security does not have an accurate price as it has not yet received a bar of data. Before placing a trade (or using SetHoldings) warm up your algorithm with SetWarmup, or use slice.Contains(symbol) to confirm the Slice object has price before using the data. Data does not necessarily all arrive at the same time so your algorithm should confirm the data is ready before using it. In live trading this can mean you do not have an active subscription to the asset class you're trying to trade. If using custom data make sure you've set the 'Value' property.

2. I removed a few further pairs from the universe (GBPUSD and EURUSD pairs as I felt these are not suitable for what the algorithm intends to do) as well as any pairs with BTC and ETH (to see how small cap coins would fare without these two),

matches = ["BULL", "BEAR", "HEDGE", "HALF", "SPELL", "USDCUSD", "USDTUSD", "DAIUSD","BTC","ETH","GBP","EUR"]

 but noticed that the BTCUSD pair is immediately added back into to Universe during the warm up period despite not being in self.tickers. The algorithm would then start trading for a while with a universe only consisting of BTCUSD

924|11:45:04:Launching analysis for 612beeb3625715aea8f8bd86c0e57b92 with LEAN Engine v2.5.0.0.14632
925|11:45:04:Algorithm starting warm up...
926|11:45:04:Added BTCUSD to the universe
927|11:45:04:Processing algorithm warm-up request 2%...

So I played around in DailyRebalance a little bit and found that using self.UniverseManager.Values[1].Members.get_Keys() instead of both self.ActiveSecurities and self.Portfolio.Keys to get the active members of the Universe seems to remedy both these issues - the algorithm won't use a symbol before it's got the price feed and BTCUSD won't enter the universe despite being blacklisted.

            universe_members = self.UniverseManager.Values[1].Members.get_Keys()
           
            ...

            weight = 1.0/len(universe_members)
            
            ...

            for symbol in universe_members:

I'm sure that I'm probably misunderstanding something and my lack of knowledge of QC's Universes surely doesn't help.

Do the changes I've made make sense?

Amended version attached.

Hi Lars,

BTCUSD pair is immediately added back into to Universe during the warm up period despite not being in self.tickers

BTCUSD is added because the algorithm subscribes to it with this line:

self.BenchmarkCrypto = self.AddCrypto("BTCUSD", Resolution.Daily).Symbol

The algorithm would then start trading for a while with a universe only consisting of BTCUSD

The FTX Crypto universe is empty at the start of the algorithm because the algorithm starts trading in January 2020 but the FTX dataset doesn't start until February 2020. The algorithm gets the BTCUSD data at the start of the backtest because the default Crypto market of the DefaultBrokerageModel is Market.GDAX (Coinbase Pro), which has a start date of 2015.

Do the changes I've made make sense?

Yes, if we want to limit the algorithm to just trade the Crypto pairs in the universe, the changes achieve that. Alternatively, we could use the SymbolData pattern and just skip over the Symbol if it's the self.BenchmarkCrypto.

Best,
Derek Melchin

Want to invest in QuantConnect as we build the Linux of quant finance? Check out our Wefunder campaign to join the revolution.

Thanks Derek,

that makes a lot of sense. I hadn’t appreciated that the benchmark symbol would itself become part of the universe.

thanks for clearing that up.

best

Lars