Choosing a Starting Point

What your first backtest should be, how many trades it should have, and how to turn it into a strategy
Published Nov 16, 2026 | Strategy-cruncher

The most common mistake systematic traders make is trying to build the final version of a strategy in their first backtest. The first backtest isn't the answer. It's the starting point: a broad version of the idea, with very little filtering, that collects the data the Strategy Cruncher needs to find the rules.

The strategy I end up trading usually has a fraction of the starting point's trades, sometimes less than 10%.

The short version: start with a signal you understand, fenced in to stocks you'd actually trade, loose enough to give the Cruncher plenty of trades but not so loose that it's mostly noise. The stronger the signal and the better you understand it, the smaller the starting point can be.

Signals and Filters

It helps to keep two things separate:

  • The signal is an event: a new high, a gap, a break of the opening range. It's the reason a trade happens at all.
  • Filters decide which of those events you actually trade.

A starting point is a universe of stocks with an entry tactic bolted onto it. The optimization is the search for the filters: under what conditions does this signal work best?

How Many Trades?

There's no exact number, and it depends on the strategy, but most starting points land somewhere in the thousands to the low tens of thousands of trades. Something like 5,000 to 25,000 is a common neighborhood, not a rule.

The stronger the signal, and the better you understand it, the smaller your starting point can be. I have strategies that start with around 20,000 trades and one that starts with about 400. That small one works because the signal is strong and very well understood. The smaller the starting point, the more carefully you have to think about every filter you apply, because each one has fewer trades to prove itself on.

Too Many

If your starting point has more than 100,000 trades, it's almost certainly too wide, whatever the strategy and whatever the time period.

It's easy to see how it happens. Make the criteria really loose, assume nothing, and let the Cruncher tell you what matters. The problem is that loosening the criteria weakens the signal, and the signal is the reason for the trade. Weaken it enough and there's no signal left, just noise. Then the Cruncher fits rules to the noise.

A backtest of every opening range breakout is a good example. Most stocks break their opening range at some point in the day, so that isn't an event, it's something that almost always happens. It needs a filter to create a universe where it means something.

Very large backtests are also slow, and they use a lot of memory to add columns and merge the trades at the end. Fewer, better trades make every step faster.

Too Few

A starting point can also be too narrow. If you have to water a signal down to get enough trades, as you often do when backtesting a single futures contract or ETF, you end up with a weak signal and a small sample at the same time.

Small samples make any split look good. With few trades, a short span of time, or many columns compared to trades, the Cruncher will always turn something up. Two hundred columns over four hundred trades will produce an impressive split by chance.

If a stronger signal leaves you with fewer trades, add more history rather than loosening the signal. Applying the signal across all US equities instead of a single instrument is the other way to get more trades without weakening it.

Spread Over Time

Check where the trades fall. If a few years hold most of them, like the COVID period in 2020, the Cruncher will favor rules that work in those years. A signal should be relatively unusual during relatively normal markets, not something that needs a rare market to work.

Universe Filters

A universe filter fences in the stocks your strategy is willing to consider. It doesn't add edge on its own the way a predictive filter does. It removes noise before the strategy ever sees it.

A backtest with 700,000 trades will include stocks with almost no volume, stocks that on most days don't trade at all. You'd never trade them, and they shouldn't be in the starting point. Cutting them out up front means:

  • Faster backtests.
  • A cleaner starting point, with less noise in it.
  • An optimization that can focus on finding edge instead of removing noise.

Is the Signal Any Good?

The more edge a strategy has before you optimize, the better the final version will be. Optimizing a bad starting point is putting lipstick on a pig.

Your starting point should be at least a little profitable before you optimize. If it isn't, step back and see if you can make a version that is, and use that as your starting point instead.

Optimizing a strategy whose starting point is weak is possible, but then the filters are doing all the work. You have to understand every one of them much more deeply, and it's much harder to tell an edge from a curve fit.

Some ways to check:

  • Run it without a stop. Hold every trade to a timed exit and see if it makes money. A stop can make a weak entry look profitable. See Backtesting Without Stops.
  • Look at the charts. Page through charts of trades from your starting point. It doesn't take thousands of them to see how strong or weak the signal is, and you'll often see something the numbers are hiding.
  • Be suspicious of "any entry works with good risk management." A random entry with good risk management is still a random entry.

Where Good Starting Points Come From

The best starting points come from intuition and experience, not data mining. Discretionary traders have an advantage here: they've watched signals play out and already have a sense of which ones have something there.

A good place to look is next to what you already trade: the same universe, or trades just outside your current rules. You'll know the setup well enough to judge it, and to believe in the strategy you build from it. That belief is what lets you trade a strategy at size through its drawdowns, and it's hard to have in a signal a machine found without a reason behind it.

Good starting points also take practice. You'll get better at choosing them the more you do it.

When a Filter Becomes a New Starting Point

Sometimes one filter changes a strategy so much that it feels like a different starting point. When that happens, treat it as one:

  • Build a new starting point. Apply the filter, then loosen some of the other rules in your original version. You'll often end up with more trades and a higher average profit per trade than before.
  • Test the filter with other signals. If the filter works just as well with similar entries, the filter, not your signal, may be the real edge. That's still useful: you've found a universe worth trading, and now you're looking for the best entry in it.
  • Look at the trades it removed. They have something in common. Sometimes they're a strategy of their own, traded the other way.

From Starting Point to Strategy

There's no workflow that fits every strategy, and no recipe like "apply two rules at aggression factor 10." A few principles hold across most of them:

  • One new rule per crunch. Every rule on a report was measured on its own, so the report can't tell you how two of them work together. Apply one, crunch again, and see what's left.
  • Match the aggression factor to the strategy. A starting point that loses money, with lots of trades, can afford a high aggression factor (33 is the default) to cut large groups of poor trades. Once the strategy makes money, a lower one (5 to 20, often with Keep Most Trades) trims smaller groups while keeping most of the trades and their profit. See Crunch Settings and Aggression Factor.
  • Every rule needs a reason. Before you add a rule, you should be able to explain why it makes this signal work better. "The data says so" isn't enough.
  • There's no fixed number of rules. You're near the end of what crunching the same data can tell you when each pass finds smaller improvements than the last, most candidates score low, each new rule leaves only a few trades behind, or the rules on offer no longer have a reason you can explain. Past that point, more rules mostly fit noise.
  • Then watch it live. Paper trade the strategy and watch the trades come in. What you notice becomes new columns to add to the backtest, and the next round of crunching starts from there.

If you've connected Claude to the Strategy Cruncher, it can read each report and suggest settings for your next crunch based on where your strategy is.