Imagine you have eight cups on a table, one tiny bell, and a friend with suspiciously good sleight of hand.
They hide the bell under one cup. You tap a cup. If it rings, you win.
That game is hard enough. Now make it stranger: instead of one cup, you have thousands. Each cup is a different strategy setting. One cup might be fast_length=12, slow_length=26, stop=2.0. Another might be fast_length=13, slow_length=28, stop=1.8. You test them all, sort by return, and the top row looks brilliant.
But here is the uncomfortable question: did you find a bell, or did you find a trick?
Atlas is built around that question. It is not trying to crown one perfect setting. It is trying to show the terrain around the setting: whether strong results form a broad, forgiving region, or whether the winner is a lonely spike surrounded by weak neighbors.
The First Principles Problem
A batch backtest turns a strategy into a surface.
The inputs are the coordinates: moving average lengths, thresholds, lookbacks, stop sizes, profit targets, time filters, and anything else the strategy can vary.
The outputs are the heights: return, drawdown, win rate, trade count, final equity, profit factor, or whatever metric we care about.
If the strategy has two inputs, you can imagine this as a landscape. One axis is the first input, one axis is the second input, and the height is performance. Good settings rise like hills. Bad settings sink into valleys.
Most real parameter sweeps are not two-dimensional. They are higher-dimensional response surfaces, which is a formal way of saying: the answer is shaped by many knobs at once.
That is why a sorted table is useful but incomplete. A table can tell you which row won. It cannot tell you whether the winning row lives on a plateau, a ridge, or the tip of a needle.
Why A Peak Can Be Dangerous
Suppose a sweep returns these three nearby settings:
| Fast length | Slow length | Return | Max drawdown | Trade count |
|---|---|---|---|---|
| 12 | 26 | 81% | 18% | 43 |
| 13 | 26 | 12% | 37% | 41 |
| 12 | 27 | -8% | 42% | 39 |
| 13 | 27 | 4% | 35% | 40 |
The first row is exciting. The neighborhood is not.
That pattern can happen for real reasons. Maybe the market had one sequence where a very specific timing rule caught a very specific move. Maybe the exit aligned with a single historical event. Maybe transaction costs, slippage, or the next month of data would knock it over.
The important point is not that the top row is fake. The important point is that the evidence is thin. If nearby settings collapse, then the strategy may be fragile. Tiny changes in the rule produce large changes in outcome.
Atlas looks for the opposite pattern:
| Fast length | Slow length | Return | Max drawdown | Trade count |
|---|---|---|---|---|
| 48 | 92 | 42% | 19% | 31 |
| 50 | 94 | 39% | 17% | 30 |
| 52 | 96 | 44% | 20% | 32 |
| 54 | 98 | 37% | 18% | 29 |
None of these rows may be the absolute winner. But as a group, they say something stronger: this area of the parameter space behaves similarly. If the exact input changes a little, the strategy does not immediately lose its footing.
That is the idea behind robust optimization.
What Atlas Means By A Region
Atlas groups nearby parameter values into bins. A bin is just a range. If a moving average length was tested from 10 to 100, Atlas might group those values into bands such as 10-20, 20-30, and so on.
Once the parameter space is binned, a region becomes a neighborhood: a combination of parameter ranges that contains many tested rows.
For each region, Atlas can summarize evidence:
- How many tested rows landed there.
- What share cleared the selected quality threshold.
- How strong the average or median result was.
- Whether the region has enough support to be trusted.
This is local aggregation. Instead of rendering every point and asking your eyes to do all the work, the interface summarizes neighborhoods.
The key word is local. A whole batch average can hide too much. A single top row can reveal too little. A local neighborhood often lands in the useful middle: specific enough to guide research, broad enough to resist one-off luck.
The Quality Threshold Is The Doorway
Atlas needs a definition of “useful” before it can score a region.
That definition might be simple:
- Return above 20%.
- Max drawdown below 25%.
- At least 30 trades.
Or it might be stricter:
- Return above buy and hold.
- Drawdown below buy and hold.
- Win rate above 45%.
- Enough trades to avoid a one-trade miracle.
The exact threshold depends on the research question. A trend-following strategy may have a lower win rate but need a strong profit factor. A mean-reversion strategy may need many trades and controlled drawdown. A strategy tested on a low-volatility market may need different return expectations from one tested on crypto.
Atlas does not make that judgment universal. It gives the researcher a way to ask: under my chosen definition of useful, which neighborhoods remain alive?
Sensitivity Screening: Choosing The Dimensions That Matter
There is one more problem. If a batch has ten parameters, not all ten deserve equal attention.
Some inputs may barely move the result. Others may control everything. Atlas uses sensitivity screening to prioritize dimensions that meaningfully change the score.
In plain English: which knobs seem to matter?
If changing a threshold from 0.5 to 0.6 barely affects performance, that input may not be worth a region map right now. If changing a trend filter length from 100 to 250 transforms the results, that dimension deserves attention.
This matters because a region view should be inspectable. The goal is not to impress the user with a cube of impossible geometry. The goal is to reduce a large sweep into a handful of places worth opening.
What A Good Atlas Finding Feels Like
A useful Atlas result often sounds less like this:
“The best setting was 12, 26, 9.”
And more like this:
“The strategy performs best when the fast length is in the 48-56 range, the slow length is in the 90-104 range, and the stop is not too tight. The region has strong support, most rows clear the threshold, and the nearby settings have similar trade counts.”
That second statement is not as tidy. It is much more valuable.
It gives a trader something to inspect. They can open representative rows, compare trade histories, test neighboring markets, and decide whether the region survives out-of-sample checks.
It also protects against a common research mistake: falling in love with the champion row before asking whether the champion has any friends.
The Honest Job Of Atlas
Atlas is not a profitability machine. It does not turn weak strategy logic into an edge.
Its job is more careful: to show where the parameter search has structure.
If a sweep is mostly noise, Atlas should make that easier to see. If a strategy has a broad plateau, Atlas should bring that region forward. If a top result is brilliant but isolated, Atlas should make the isolation obvious.
That is the deeper capability: moving from row-picking to terrain-reading.
When you analyze a batch of test results, the first prize is not the highest number. The first prize is understanding what kind of landscape produced that number.