The Filter(s)

In theory you should be able to build a breakout strategy with just a good POI, the right FRACT in the SPACE calculation, and a reasonable stop-loss and profit-target.

In practice, these over-simplified models never work well.

In my experience, strategies without a filter always failed very quickly and lost money.

Why do breakout strategies always need a filter?

I was once intrigued by the simplicity of the “POI and SPACE only” approach. I could even run beautiful 3D visual optimizations on it. It became obvious very fast that the formula always needs a filter.

And when I say always, I mean it. Four reasons:

1. A simple model without a filter produces too many false breakouts.
2. It produces way too many trades.
3. Because of those two, the Average Trade is almost always too low to call it a viable strategy.
4. Breakout strategies without a filter simply do not perform well in real out-of-sample (live) trading.

Oversimplification doesn’t work in trading. There always has to be a reasonable balance.

Which filter should you use?

That’s the wrong question.

With filters, exactly as with POIs, you need to try many different approaches to find what works for each market and each timeframe.

The much more important question is not WHAT.

It’s HOW MANY.

It took me a long time to figure this out. Because the easiest thing in the world is to stack filters on top of each other. It looks beautiful on historical data. What you actually end up with is a completely overfitted, curve-fitted strategy that fails from day one in your live account.

Overfitting and over-optimization is the number one mistake beginner algorithmic traders commit. Again and again. Obsessed with easy money, they fine-tune their backtest reports so tightly there’s no room left for even the slightest possibility of survival in real trading.

I committed this crime myself. I used to build strategies with four or five filters, get fooled by incredible backtest equity curves, and then lose money at the speed of light live.

At the peak of my frustration I removed all filters and tested pure breakout models. That failed too.

So I went back to filters, but reduced to one or two. Better. Still uncertain. I needed a quantifiable answer.

What did our hedge fund study find?

In the early days of my hedge fund we ran an extensive study to finally solve the “how many” question.

At the time we had over 700 different breakout trading strategies, mostly futures, covering intraday and swing timeframes. All had passed our robustness tests and met our quality requirements for live trading.

Every strategy in our database is monitored constantly, which gives us real out-of-sample trade data. When we ran the study we had about 18 months of true out-of-sample performance — enough for reliable conclusions.

We took that real out-of-sample data and split it into seven groups according to the number of OPTIMIZATION INPUTS. We had strategies with anywhere from one to seven inputs.

For the assessment criterion we used our proprietary index, avgDhidx. It’s like the TradeStation Index but more complex: it measures quality based on Net Profit vs Drawdown performance plus the quality of the equity curve — its stability and its correlation to an “ideal” curve. Higher avgDhidx means a smoother, less bumpy curve and a high Net Profit / Max Drawdown ratio.

We calculated avgDhidx for each strategy’s true out-of-sample performance, then averaged by group.

The study wasn’t perfect. Perfection isn’t our standard — we learned from Jeff Bezos that 80% of optimal is enough to make an assessment and move forward faster. If it’s good enough for Amazon, it’s good enough for us.

Three conclusions:

1. We could find no evidence that simpler strategies are better. None. This confirmed that oversimplification is not the solution.
2. More than six optimization inputs could potentially be dangerous — but between two and six inputs there was no significant difference in true out-of-sample performance.
3. The group with four inputs scored oddly low. It was also the smallest sample. My gut says it would normalize higher with more data.

Conclusion: with strategies built to my formula, you can confidently run five or six optimization inputs and still be robust.

How many filters does that leave you?

Let’s count.

We always optimize FRACT in the SPACE component. That’s input one. Then you’ll likely optimize a stop-loss and a profit-target. That’s three.

Six is the maximum. So you have three slots left for filters.

It really doesn’t matter which filter you use. What matters is keeping the number of optimization inputs within range.

For example, use RSI and optimize two parameters — period and threshold — and you have five inputs total:

INPUTS:
Fract(1),
StopLoss(500),
ProfitTarget(1000),
RSI_filter_Period(25),
RSI_Threshold(20);

Filter_Long  = RSI(RSI_filter_Period) > (100 - RSI_Threshold);
Filter_Short = RSI(RSI_filter_Period) < (0 + RSI_Threshold);

You still have one input left, so you could add a second filter:

INPUTS:
Fract(1),
StopLoss(500),
ProfitTarget(1000),
RSI_filter_Period(25),
RSI_Threshold(20),
MA_Period(100);

Filter_Long_2  = Close > MovingAverage(MA_Period);
Filter_Short_2 = Close < MovingAverage(MA_Period);

Now stop reading for five minutes and write down at least five filter ideas of your own. Creativity is the key here.

What filter groups do we use?

In the hedge fund we don’t think in individual filters. We think in filter groups:

• Price action-based filters
• Volatility-based filters
• Trend strength-based filters
• Volume-based filters
• Moving averages-based filters
• Experimental (proprietary) filters

When we build a new strategy we test the groups first — each contains many filters. Once we see which group responds best for that market and timeframe, we narrow down. Then, within the input budget, we experiment with combinations of up to two filters, preferably from different groups: one from the best-responding group, one from the next best.

We don’t do this manually anymore. My team built sophisticated automation for it.

Think of the “how many” guidance as a formula within the formula.

Next component: TIME.