Source: Deep Learning on Medium
How I Made An AI Trade 100 BTC To 120.47 BTC In 76 Days.
And why a trading firm in Israel is giving my AI a “go” to trade their funds.
“100 BTC to 120.47 BTC within 3 months. Your performance is good and I hope you will continue with the same pattern!”
I have written a deep learning model that trades crypto. Yes cryptocurrencies. Today, Alefbit Technologies Ltd, a boutique Israeli crypto trading firm, wants it deployed.
In 76 days, ADL (Amrullah Deep Liquidity), the trading model I have developed, returns a 20.47% profit on Alefbit. No use of leverage. The model is put through a live audition process on Alefbit’s proprietary platform.
Let’s analyze the trading metrics which are no less remarkable.
Trading Period: 15-July to 15-Oct
Profit factor: 3.32 (1.5 is good, 2.5 is excellent, above 3.5 is world class)
“If I spend 1 bitcoin I can expect to get 3.32 bitcoins back from trading this model.”
Gain-to-pain ratio: 2.32 (1.0 is good, 2.0 is excellent, above 3.0 is world class)
“I expect to handle a small loss to gain 2.32 times more”
Sharpe ratio: 15.65 (1.0 is acceptable, 2.0 is excellent, above 3.0 is top)
“I expect to make 17.29% — 19.65% return every quarter, 68% of the time (which represents one standard deviation)”
Max drawdown: 4.39%
“The max I lose from any all-time-high is 4.39% from trading this model”
A month ago I have been featured in a documentary called “Coded World”. I show how to build high-performance facial recognition systems using deep learning techniques. This time, I show how to use deep learning to trade the cryptocurrency markets.
ADL is built to identify market structures left by other market making bots. It uses deep learning to achieve an unprecedented level of accuracy.
The model works by guessing the inventories of other market making bots. The goal is to detect any imbalance in these inventories. These bots are more likely to skew the order book to restore the balance.
The ultimate test for a quant model like ADL is to enable it to take positions in the market with this knowledge. And so, ADL did!
76 days and 64 trades later, ADL makes a healthy rate-of-return. The growth of the equity curve is steady and the days under are short.
The Crypto Firm
And so how does Alefbit enable quant traders like me to thrive and trade their funds?
Alefbit hires talented traders around the world. The traders trade to generate outsized quarterly returns for the firm. “At least 20% per quarter!” That is the impression I get during the audition for the firm.
Potential traders go thru a 3-month grueling process. This grueling process tests their trading abilities. Once successful, they get remotely hired to trade on behalf of Alefbit.
As a 28-year-old data scientist who designs quant models, I’m one of many that has auditioned for the firm. The firm claims to have “more than 200 traders at different stages of their trading trial period”.
“Quant models can generate outsized risk-adjusted returns. It uses math to place the right trades at the right time with the right size.”
Today, the firm is planning to use ADL’s quant model by allocating funds for me to trade with. As ADL performs, the firm allocates more! Simple? Not so.
The boutique trading firm cares a lot about keeping drawdowns low. This fact is a far cry from the risk profile of other crypto hedge funds. Some posts quarterly drawdowns of up to 22.92% for the same period last year!
There is no magic number. After speaking to the Head of Trading at Alefbit, Ilan Sterk, I have a good gauge. The sense I get is to adjust the model’s max drawdown to under 5%.
Backtesting and Customization
Quant models allow users to tweak their initial settings and parameters. Users can then adjust the theoretical max drawdown to levels that they need.
The whole process of selecting the initial parameters of ADL can be fun. It enables traders to tweak and experiment the models to suit their trading style.
Lo and behold, it works like a charm in reality. It posts an actual max drawdown of 4.39% throughout the 90 days audition.
In the case of Alefbit, the firm has strict risk management protocols. The protocols limit the tradable assets, limit exposure to a single asset and set a cut-loss. The only way ADL can pass through the audition is (1) to return a smooth equity curve (2) to keep a profit factor that is greater than 3. In other words, I have to tweak ADL’s initial settings to focus on SHARPE RATIO!
“The Sharpe ratio can be used to compare risk directly between two funds. The comparison shows how much risk each fund has to bear to earn excess return over the risk-free rate.”
The trade-off of increasing Sharpe Ratio is Profit Factor.
“The Profit Factor is a better metric to reflect the projected ratio of win-size/loss-size.”
The experiment is to find the sweet spot on the curve that can generate the returns you aim for. Voila!
ADL manages to keep a low standard deviation of 1.18%. This translates to a Sharpe Ratio of above 15.65 while keeping a Profit Factor of 3.32.
Using ADL shows how quant models are versatile tools to trade the crypto markets.
If you have what it takes to trade for Alefbit, you can use my referral link here: https://alefbitech.com/?wpf249_4=0BA5080AF6
If you run a crypto fund or a trading shop, you’ve come to the right place. You can check my TradingView to test out ADL’s capabilities here: https://www.tradingview.com/script/PvF8Ai47-Amrullah-Deep-Liquidity/
You can also reach out to me on Twitter: @Muhd_Amrullah
A special thank you to Linh for graciously supporting me through thick and thin.