# What are the best metrics to evaluate an ML algo backtest, other than the nominal returns?

I have designed an algorithm that uses Support Vector Machines to classify the next day's price movement for several prominent cryptocurrencies on a [1,0,-1] (buy/hold/sell) basis. These cryptocurrencies are namely Bitcoin, Ethereum & Litecoin. I have desgined a backtest that takes into account fees, spread and depth too and therefore can assume a backtesting environment somewhat more realistic than most. When aggregating the returns from across these 3 cryptocurrencies, the overall return is positive.

But there must be other, more insightful and prudent metrics that indicate the success of a strategy other than just pure returns. The only metric that I have so far compared the strategy to is a Buy and hold (B&H) and comparing returns to the CCi30 index, an index for cryptocurrencies. Whilst I have seen a similar answer here, I found the answers slightly vague and not necessarily relevant to this particular asset class.

Are there any other metrics that can evaluate the performance of a backtest? And which ones should receive the most weighting?

• Average performance over index seems to be a good one, make sure you're taking the geometric mean over time as well. You could also find the volatility if your strategy and from that the sharpe ratio, max drawdowns woukd be interesting as well as would the win rate May 6 '20 at 3:52
• If you try to "sell" this strategy to investors they will want to know the volatility (or tendency of returns to fluctuate from day to day) as well. So standard deviation of returns is helpful to calculate. May 6 '20 at 14:15
• Thanks for your comments Oscar, your points about geometric mean seem interesting and relevant. Could you elaborate on what you mean by volatility and the win rate? Thanks. May 8 '20 at 15:27
• If you back test your strategy 5 years back you're going to get a value for your hypothetical portfolio for each day of those 5 years. Starting at some value, fluctuating a bunch, and ending (hopefully) at above the start value. Simple take the sample standard deviation of the returns from that time series. Regarding win rate I mean when your strategy buys and later sells a position, what is the probability it made a profitable trade? Note that you could have a profitable system with a low win rate if you have few big winners and many small losers, but it might be a less attractive system May 9 '20 at 4:41