Almost all of us think that MORE IS BETTER. That the more indicators, conditions, and online trading theories a chart has, the more likely it is to win.
This is exactly the kind of thinking we should avoid at all costs. Many of us blindly believe that the more confirmations we have, THE BETTER it is for our trading.
That’s how I spent an entire year using a strategy with a good trigger that ended up being OVERFITTED because it had so many complementary indicators. And that’s exactly what I’m going to talk about in this last post in the series on How to Do Professional Backtesting from Scratch: (Why Your Strategy Will Fail Live) and (From Idea to Algorithm: Which Platforms and Data to Use).
So, in trading, what matters is not the number of indicators and confirmations, but the hypothesis underpinning your strategy: what logic it is based on—or, more simply, WHY SHOULD YOUR STRATEGY WORK? If you can answer that and back it up with historical data, and the results are positive, then you have a strategy based not on opinions, faith, or street logic, but on certainty that comes from validating your idea with data.
Let’s get into it: How can I tell if my strategy is overfit, and how do I validate it? In other words, how much HISTORICAL DATA do I need for it to be considered consistent?
According to Perry Kaufman (Guide to Creating a Successful Algorithmic Trading Strategy), one of the living legends of systematic trading, consistency does not depend on the number of years of data, but on the size of the sample of trades (Sample Size). In his view, evaluating fewer than 100 or 200 trades produces erratic and statistically irrelevant results.
So the idea that the more years of data I cover, the more market regimes I can capture is counterproductive. It would be like setting ourselves the goal of finding the Holy Grail of trading—a strategy that works in every regime. It’s better to focus on building a portfolio of specialized systems, each of which excels in a specific regime.
On the other hand, as for OVERFITTING, as we’ve noted, it happens when we force our strategy to meet too many conditions at once—for example, to trigger an entry, take profits, or keep a trade open.
In backtesting, this means forcing the algorithm to meet too many conditions in the past to “increase profits” and make the strategy look perfect, with no losses—which is A FLAGRANT TRAP.
You haven’t created a winning system; you’ve just created code that memorized the answers to the questions the teacher asked yesterday.
When the market changes its personality (a regime shift whose timing you can’t predict) in real time, your account will suffer the devastating consequences of overfitting.
In this regard, Roberto Pardo, author of Walk-Forward Analysis, points out that since backtesting is a stress simulation, a sufficiently robust system is one that knows how to perform in its ideal regime and has risk-management rules to minimize losses when the market enters a hostile regime.
Thus, expecting a single strategy to be perfect in every scenario is a mathematical utopia that only leads to overfitting. A professional isolates the data to determine exactly which specific regime gives their strategy an edge, and designs robust risk management to protect capital when the market changes its personality.
If you want to conduct a statistically rigorous backtest, here are a few rules you might consider:
1. Forget indicators; evaluate time: A consistent system doesn’t depend on miracle arrows, but on raw statistics. For fast strategies such as scalping or day trading, you need to test your logical rules on at least 3 to 5 years of pure historical data (sourced from reliable providers). This forces your robot to face every possible regime: bull markets, brutal crashes, and sideways liquidation ranges. If it only wins when all the indicators are perfectly aligned, as if the stars were aligned, it’s not a strategy—it’s a lottery ticket.
2. The Blind Test: To find out whether your strategy is real or just a visual illusion, apply the institutional rule: divide your data. Use the first x years to design your robot’s rules. Then let it trade over the remaining x years without changing a single parameter. If performance drops sharply or it starts losing during this final period, your system was overfit. Discard it without hesitation.
3. Execution: The final adjustment to make your simulation identical to reality is to program your trigger using the Next Bar Open protocol. The algorithm analyzes the candle close and calculates the edge, but executes the order at the exact microsecond the next candle opens (00:00 UTC). Trading in the session’s interim noise destroys your edge.
Systematic and quantitative trading does not require memorizing candlestick patterns, filling your screen with colored lines, or adding a hundred thousand indicators that give you conditions for entering the market, exiting, or keeping a trade open.
It requires accepting that less is more and that the only guarantee of survival is a proven statistical edge.
These are reading times.
#EarningsSeason #TradingSistematico
