Excerpt from a conversation with a client about backtesting and how to approach quant trading.

People tell max dd numbers to their client like they know what is going to happen in future. This is totally wrong and tantamount to misleading the client.

We never believe in telling expected MDD numbers to clients because there is no way to really know that number.

Some are saying take 2x MDD as future MDD. But why 2x? Why not 1x, 1.5x, 3x, 5x, 10x?? It seems like totally arbitrary choice.

They say they will decide based on how good backtest is done but there is no quantified way to tell how well backtest has been done.

We have made so many algos. Few stay in line with out of sample performance, some totally outperform, others totally underperform.

There is no way to know which is going to do what. Only time in markets can tell that information. So keep note that backtesting is not at all something that tells you what future performance will be.

It only tells whether your algo has any probably positive expectancy or not.

On that note, if anyone tells you their strategy has positive expectancy then they are lying or don't understand ergodicity and the lack thereof.

Positive expectancy of algo is not fixed thing. It is a stochastic property - it will change with time and you can never be sure.

So we do not worry about positive expectancy. We are in game of applying law of large numbers to "probably positive expectancy" algos.

If you make many uncorrelated probably positive expectancy bets, you are likely to realize that positive expectancy

In same way, each algo is a bet for us. This is a game of probabilities and that's how we like to approach it.

We are not trying to make money for one day or one week or one month. We want our clients to have profitable portfolio in long run.

Decent cagr at scale is much better for wealth building than big returns that have high risk of ruin or that don't scale.

Buffett is perfect example. His wealth is result of ~20% cagr. But reason he is billionaire is because he can create that return on very big portfolio also.

Make your approach of probabilities. Use LLN on not just trades, but algos.

Allocate money to 5 different algos each with some MTM buffer (like in chem burn it is 10k MTM buffer out of 50k). Then let them play out. Stick to the system and let it do the trading.

When any algo runs out of its mtm buffer, evaluate if it's still worth betting on or is there better algo available.

If you do this repeatedly, in long run nobody can deny you a profitable portfolio.