L/S strategy on Bybit using Numerai Crypto Metamodel
前のエントリの英語版です
Introduction
I’m dera. I usually work as a postdoctoral researcher in a university lab.
In this article, I’ll do a rough verification of what would happen if we actually used the Numerai Crypto Metamodel predictions for trading. (You can follow along even if you’re not familiar with Numerai.)
What is Numerai Crypto?
Numerai is a platform where you stake the cryptocurrency NMR to predict the stock market four weeks later. Depending on your prediction results, you can mine NMR. Numerai’s management aggregates all submitted predictions proportional to their stake amount (= Metamodel) and uses that Metamodel to operate their hedge fund. According to the September 30, 2024 report, the fund’s AUM is $375M. There was a period of large losses for a while, but they’ve been doing well lately (quoted from Kei Sanada’s blog).
For more details on Numerai in Japanese, the following article by UKI is well-known:
The primary targets of Numerai’s predictions are global stocks. However, in May of this year, a trial version of Numerai Crypto was launched to predict cryptocurrencies. Unlike the existing stock market prediction services (Tournament and Signals), Numerai Crypto’s Metamodel is published in real time.
In other words, we can actually refer to the predictions submitted by NMR stakers and use them for trading. The performance of the Metamodel is shown as the rank correlation coefficient (CORR) between the predicted 30-day returns and the actual returns of the target. It looks quite impressive with an upward trend. A high CORR means that the 30-day returns of each target cryptocurrency are accurately ranked by the predictions, so it seems like we could make a profit by trading based on these predictions.
Therefore, with the idea of trading on a CEX (bybit), I performed a backtest based on PnL (rather than CORR) using the Metamodel’s predictions.
Basic Trading Rules
The Metamodel’s predictions are given as numerical values ranging from 0 to 1, like in the image below.
For Numerai’s fund operation, they use a market-neutral strategy that goes long on the top X-ranked assets and shorts the bottom X-ranked assets based on Metamodel predictions. We will follow that same approach this time. For example, if we only had the above chart of assets and X=3, we’d go long on BTC, ETH, and BNB and short DOGE, XRP, and ADA. Since Numerai Crypto predicts the 30-day return, we hold each position for 30 days. (Strictly speaking, it’s 20D2L, i.e., 20 business days plus 2 days of lag, but for simplicity, we’ll just say 30 days here.)
Fees
I’ll assume taker fees (0.055% * 2 round trip) plus slippage (0.05% * 2; a rough, likely overestimated assumption from bybit’s order book) as trading costs. Ideally, We should also consider Funding Rates, but generally, each crypto’s funding rate moves in tandem with market hotness, so going long/short in a market-neutral fashion should make it roughly neutral overall. (Actually, it was just too much work to incorporate. Realistically, we should account for funding rates. They can’t be ignored.)
Selecting Tickers
The universe of target assets for the Metamodel predictions consists of the top 500 cryptocurrencies by market cap. LSTs and stablecoins are excluded, as shown below:
But among those 500, many are extremely low-cap “shitcoins” with no perpetual futures available, making it impossible to short them. You could use DEXs or lending platforms to short some of these, but if we can’t trade long/short in the same place with cross margin, we lose the benefit of leverage that a market-neutral strategy typically affords. So for this experiment, I’ll only consider assets that have perpetual futures on bybit.
I checked the overlap between the Metamodel’s set of 500 coins and around 400 of bybit’s actively traded coins, and found that 275 coins are eligible for our trading.
So, 275 coins have potential trading targets.
Backtest
We only have half a year’s worth of data (May–November) during which the Metamodel’s predictions are available and enough time (30 days) has passed to measure outcomes. Numerai’s actual operation supposedly goes long on 50 out of 500 tickers and shorts another 50 (I think I read that somewhere, but I could be mistaken). So here, for our 275 available tickers, I’ll just pick 25 long and 25 short for each rebalance.
Although crypto trades 24/7, the Metamodel is based on a stock market pipeline, so the predictions aren’t updated on weekends. Over the course of 30 days, we’d open/close positions on at least 20 (week)days. Hence, our margin gets split 20 ways, and the returns effectively shrink to about 1/20 per trade. Overall, the half-year return is around +50%. The Hit rate is quite high, but the drawdowns are also large, so it’s not easy to use much leverage. It seems to have captured the recent bull market, but around September, it repeatedly shorted pumping coins such as REEF, causing big losses.
Since some of those pumping coins likely had crazy funding rates at the time, if we’re holding positions for 30 days, we really should factor in the funding rates. A significant change in the PnL curve can occur by considering FR.
Conclusion
Given that the CORR chart is a flawless upward slope, the results here are a bit disappointing. Because there are many coins that we can’t short, the Metamodel’s current rankings are not very convenient for real trading. In this backtest, I tested a market-neutral strategy using the Metamodel, but the crypto market at the moment is bullish, so maybe a long-only approach may be effective.
Currently, Numerai Crypto is basically repurposing the existing stock market prediction pipeline to crypto, so there are no weekend predictions, and the prediction window is relatively long (30 days). It might be more practical to shorten the timeframe or narrow down the target list if we want a Metamodel that can be truly useful in real-world trading in my opinion.
Anyway, thanks for reading. Here’s to next year!
Discussion