Strategy Selection

Using Predictions for Trades

In order to actually deploy trading bots (not only their "brains", which we like to call the models), we need financially sound investment strategies. Optimally trading on a prediction is no easy task and can be generally described as bet-sizing.

Bet-sizing is a whole subfield of financial mathematics, but generally, we aim to maximize expected returns, given the expected (backtested) probabilities to correctly predict the price direction of an asset in some time frame, which has been elucidated in the previous chapter.

There are of course other things we can predict, like volatility, and use them in trading, for example with options. The methods differ, depending on which assets or derivatives one wished to make a market for.

Example: Leveraged Spot Market

Take the following example: A market like Mango provides cross margin accounts to its users, allowing them to trade certain assets with up to 5x leverage given the invested funds. This also allows them to borrow those assets and sell them - shorting them. This gives us multiple options:

• Limit the leverage and thus the risk we are willing to expose ourselves to.

• Limit the trade directions, allowing only for long, short, or both kinds of trades.

Limiting the direction allows us to strategically employ the predictions of the algorithm, in order to reflect the trader's opinion of the current market situation:

• Allow only shorts, when they expect the market to correct or become bearish

• Allow only longs when there is a bullish outlook

• And allow both if the trader has high confidence in the predictive abilities of the model.

Limiting leverage may be appropriate when the stability of returns is not guaranteed, given a low Sharpe ratio, which will be explained in the next section.

Financial Metrics

In order to formulate a financially optimal strategy given our model validation results, it will be necessary to define some metrics which describe the performance of the model more accurately.

The Sharpe ratio

Depending on the reliability of the predictions, we might want to limit leverage i case the algorithm maye eve events of repeated large losses and repeated strong gains. Another way to look at this kind of behavior is in the context of the Sharpe ratio. The higher it is, the higher the stability of returns, which allows us to increase leverage. A high Sharpe ratio for the projected returns of our model allows us to increase leverage in order to increase the mean returns. Meaning we can bet more money on the predictions of our model without risking too much of our funds.

The Portfolio Delta

The delta of a portfolio describes how much the portfolio's value changes, given changes in the price of held assets. Given that some assets historically correlate strongly with each other, it is possible to short one and long the other just as much, creating a zero delta, or market-neutral, position. This can be useful to limit our portfolio risk to the crypto markets' ups and downs.

On the other hand, this means that a market-neutral strategy purely derives its profits from the model's ability to correctly predict price moves!

If our model detects such a strong correlation between two assets, but its prediction sometimes breaks this correlation -longing one and shorting the other- then this means that it picked up some pattern which those two assets' prices share. This is the famed alpha our model picked up, given the data and its hidden patterns.