How To Explore a Trading Idea

In my last video, I explained how generally speaking, trading strategies shared publicly don't work because of alpha decay. If a trading strategy has a legitimate edge, then traders will pile in and crowd the trade until the edge no longer exists. A comment on Instagram pushed back by asking me to test this order block-based strategy. So, instead of just arguing in the comments, I think this is a good opportunity to show what testing a trading idea actually looks like very briefly. Not going to open up a TradingView replay session, backtest this over 25 or 50 trades, maybe see a green P&L, and call it a day and say "Yes, this works. Let's go deploy this live and risk real money with it." No, we're going to program this. We're going to backtest it programmatically over half a decade worth of market data, then we're going to statistically verify whether it is indeed distinguishable from random noise. The strategy was very simple. It takes place in regular trading hours only. We mark out the first 15-minute bullish order block, and then on the 1-minute time frame tap of that order block, we enter a long trade targeting a fixed multiple of risk with the stop below the bottom of the order block. Testing this bare-bones version of the strategy over 5 years of Nasdaq front month contract data, we get these results here. It doesn't look like it completely collapses at first glance, right? We have an equity curve that goes up into the right somewhat, right? We end with more money than we started with. But, that's not the end of the story. That's not the full story at all. When we bootstrap the 95% confidence interval for trade return by day, we find out that it spans zero. It spans negative numbers. We can't say with a high degree of confidence that this strategy actually makes money over very many trades. It's not statistically clean enough to call a verified edge. What's interesting though is that if you compare the actual order block timing versus a randomized entry model after the signal is formed, a null model, then the actual order block timing beats the null model with very low P values. We have a P value around 0.01 for mean R and essentially zero for net P&L. The strategy also performs much better in our test set, which starts around 2025 than it did in the training set, which shows that for whatever reason, it seems to be working much better now than it was a year and a half ago. Experimenting with the entry logic a little bit tells us that a tap of the top of the box isn't actually the optimal time to enter this trade. If we instead wait for price to move down to the bottom 25% of the box, our entries get much better. Our mean net R per trade becomes a lot higher, especially with a 1.5 R target. So, let's look at what that looks like. Just to be clear, obviously to avoid over-fitting, it was a training set that we came up with these parameters on. Now, let's look at it out of sample. test set, this slightly optimized version of the strategy produced 199 trades for a 0.339 mean R per trade. Bootstrapping the confidence intervals again, this time we can say that they are fully positive and they don't span through zero. The optimized version of the strategy could potentially be a valid jumping off point for a prop firm strategy, but it drastically underperforms buy and hold in the real world if you were working with live capital. So, just like that, it only took us a few minutes to start exploring and probing this trading idea properly. And in those few minutes, we've already made a major discovery about a promising optimization. This is how it's supposed to start, with an idea, an exploration, and optimization, not watching YouTube