Historical data rejects. It does not approve.
A good backtest only suggests the strategy is worth further testing.
Trading strategy mining
Out of a million strategies, almost none are worth your money.
Your agents generate millions of strategies and test them on historical market data. A statistical filter discards almost all. The few that remain operate for months on paper, with real prices. Here you see the entire process, step by step.
The SieveEJEMPLO
- EJEMPLO10,000,000 candidates
- EJEMPLO100 pass to validation
- EJEMPLO3 to incubation
- EJEMPLO1 in the room
The Idea, in Three Sentences
What matters is what happens next.
The trades a strategy makes from its discovery, on paper and with real prices.
That is slow, and cannot be bought with computing power.
It is measured in trades, not days. More computers do not speed it up.
How it works, step by step
Six steps, from assignment to the trading floor.
01
You choose an asset and hire an agent
An agent is a search process dedicated to an asset and a timeframe: for example, BTC on 1-hour candles. It does not predict the market and does not trade on its own: it searches for rules and puts them to the test.
EJEMPLO
Agent · BTC · 1h
searching
02
The agent builds strategies with simple components
Each strategy is a short rule: one or two entry conditions, one exit, a stop, a target, and a maximum duration. The components come from a closed catalog of 24 known indicators (moving averages, RSI, channels, MACD…). Combining them yields over a hundred billion different rules.
A RuleEJEMPLO
- Entry 1
- close > EMA 200close crosses above SMA 50EMA 20 > EMA 50close > EMA 200
- Entry 2
- RSI 14 < 30ADX 14 > 25MACD hist. > 0RSI 14 < 30
- Exit
- RSI 14 > 55close < EMA 20no conditionRSI 14 > 55
- Stop
- 2 ATR1.5 ATR3 ATR2 ATR
- Target
- 3 R2 Rno target3 R
- Maximum time
- 96 candles24 candlesno limit96 candles
03
It tests them on historical market data, but not on all of it
The historical market data is split into three segments by date. The agent only sees the oldest segment. The other two remain on the server: one for filtering and another that is looked at only once, at the very end. This way, a rule cannot memorize the data it will later be tested on.
55 %
30 %
15 %
oldestnewest
- Mining55 % · what the agent sees
- Validation30 % · server only
- Reserve15 % · single glance
04
The filter discards what was only lucky
By testing millions of combinations, some will always appear that would have won by chance. That's why the bar rises with each strategy tested: the more tests, the more a strategy must prove. All tests are counted, including those from other campaigns on the same asset.
In technical terms: a Sharpe ratio deflated by the number of tests.
The BarEJEMPLO
strategies tested
05
The remaining ones paper trade, with real prices
Passing the filter does not grant entry to any trading floor. First, the strategy paper trades: the server follows its signals with real market prices and without real money. This is intentionally slow: a strategy on 1-hour candles might only make a few trades per month. After 30 trades, the examination arrives: if in paper trading it doesn't resemble what it promised, it is discarded.
EJEMPLO
paper trades0Examination
06
The trading floor shows their history and the ranking sorts them
On a trading floor, you see what each strategy has done in paper trading since it was found. The ranking does not use backtests: it sorts by a conservative score, which considers the minimum to expect rather than the average. Few very good trades are worth less than many regular ones.
EJEMPLO
BTC30-99
ETH30-99
Gold30-99
Everything is measured in R
1 R is what a strategy risks in one trade: the distance between the entry and the stop. If the stop is hit, the trade loses 1 R. If it reaches a 3 R target, it wins three times that amount. Commissions and slippage are subtracted from each result. Measuring in R allows comparing strategies across different assets and sizes.
EJEMPLO
A price line starts from the entry. If it reaches the target, which is three times further than the stop, the result is +3 R. If it drops to the stop, the result is -1 R.
The Noise Test
Why a backtest doesn't decide anything here
We mined a million strategies on invented prices, where there is nothing to find. In the backtest, the ten best earned +0.26 R per trade. In the unseen data, they lost 0.05 R per trade. That's why a backtest doesn't decide anything here.
R per trade of the ten best
candidates tested
- what the backtest promised
- what it did afterward
How to read it: the more candidates tested, the better the backtest of the ten best performs. What they do afterward remains slightly below zero. The distance between the two lines is luck.
Eight synthetic series with no edge, 5 years of 1-hour candles.
What's in a Room
A room is a collection of strategies with their paper trading history. From the outside, you see a summary that is updated once a week: trading bracket, approximate score, and whether it has a badge.
EJEMPLO
StrategyPaper TradesApproximate Score in RStatus
Score: the minimum expected from the strategy per trade, in R, based on what has been observed so far. It's intentionally conservative: with few trades, it may appear negative even if the strategy is performing well.
Badge: the score is already above zero based solely on its own paper trades. A good strategy needs about 200 paper trades to earn it.
With fewer than 30 paper trades, it does not enter the ranking.
One Agent Per Asset
Each agent works with one asset and one timeframe, with a limited testing budget: over-testing does not improve the result, it worsens it.
BTC · 1h
- candidates tested
- 10.000.000
- in validation
- 100
- in incubation
- 3
- in room
- 1
ETH · 4h
- candidates tested
- 8.400.000
- in validation
- 84
- in incubation
- 2
- in room
- 1
Gold · 1d
- candidates tested
- 6.200.000
- in validation
- 62
- in incubation
- 2
- in room
- 0
Your device can mine
You install a small program. It only performs calculations: it does not store exchange keys, only opens outgoing connections, and pauses on battery power. If you participate in an open room, you accumulate discount credits.
This is not a way to make money with your computer.
How work is verified as real
- 01
Your device receives a batch of candidates and calculates them.
- 02
It sends a sealed summary of its results, before knowing what it will be asked.
- 03
The server randomly selects which parts to check and recalculates them.
- 04
Only work that passes this check counts.
Your device
batch
Server
What this platform does not do
It does not predict the market.
It does not touch your money or store your exchange keys.
It does not tell you what to trade.
It does not convert a good backtest into a good strategy. Only time tells that.
Pricing
Plans differ by the number of agents and rooms.
Price on request
For now, sign-up is closed: accounts are created by the platform. Tell us about your case and we will get back to you.
Go to the contact formOffer this to your community with your name and your colors.
Glossary
- Agent
- A search process dedicated to an asset and a timeframe.
- Candidate
- A strategy generated and tested on historical market data. Almost all are discarded.
- R
- What a trade risks. All results are expressed in R.
- Paper trading
- Following strategy signals with real prices but without real money.
- Room
- A set of strategies with their paper trading history.
- Score
- The minimum expected from a strategy per trade, in R, based on what has been observed so far. It ranks the strategies.
- Badge
- The score is already above zero based only on the strategy's own paper trades.
- Credits
- Discount on your fees for the work contributed by your team. They are not money.
Frequently Asked Questions
What is an agent?
A search process dedicated to an asset. It generates strategies, tests them, and delivers those that pass the filter. It does not predict the market and does not operate on its own.
Do I have to leave the computer on?
No. You can mine on your computer or hire computation. If you mine yourself, the program pauses with battery.
Do you touch my money or my keys?
No. The platform does not connect to your exchange or store keys.
Why is almost everything discarded?
Because testing millions of combinations always turns up one that would have won by chance. The filter is for that purpose.
What are credits?
A discount on your fees for the work your team contributes. They are not money and cannot be withdrawn.
Is this advice?
We do not provide personalized recommendations or tell you what to trade. What you see are results from paper tests.