How to Choose Crypto Trading Pairs for a Bot Without Cherry Picking

Start with a rule for which pairs were eligible

A trading pair is not eligible simply because its candles can be downloaded today. State the exchange, spot or perpetual market, quote currency, minimum data history, minimum tradable size and liquidity threshold before running an optimization. The rule must use information available at each selection date.

For example, a research rule might admit USDT pairs listed on a chosen venue that had already traded for 90 days and met a predefined recent volume threshold. Those numbers are illustrative choices, not universal safe limits. Record exactly which pairs passed on each review date and when each became available.

Why picking the best backtests afterward is a trap

Suppose a bot is tested on 30 pairs and the published chart shows only the strongest three. The question has silently changed from whether the rule works on the intended universe to whether three favorable histories can be found. The more parameter combinations and pairs examined, the easier it becomes to find attractive results by chance.

A hypothetical report with 30 pairs should show the full distribution: the median, the weakest pairs, the number with positive net returns, trade counts, drawdown, time in position and the common test dates. Do not average percentage returns from independent full balance backtests and call the result a portfolio.

• Show the full original list of 30 pairs and the selection rule.
• State how many settings were tried per pair.
• Retain failed, missing and delisted markets in the audit trail.
• Separate selection period performance from untouched validation.

Account for coins that disappeared

Choosing today's active coins for an old test can omit coins that disappeared, lost liquidity or were delisted. That is survivorship bias. A cryptocurrency research paper finds that excluding such assets can materially distort historical performance measures; its numerical estimates depend on its sample and portfolio construction, so they should not be applied as a correction factor to an individual bot.

If historical listing and delisting data are incomplete, say so. Reconstruct point in time eligibility where possible, save archived pair lists and report the missing coverage. A test on today's surviving pairs can still be useful as a conditional experiment, but it cannot establish how an unchanged historical selection rule would have behaved.

Simulate one account, not thirty separate accounts

A 10% return on each of two isolated tests does not mean one account could have earned both. Signals may occur at the same time, capital can already be committed, and exchange risk limits or a maximum position count can block a new order. A portfolio backtest should replay all eligible pair signals chronologically with the same cash, sizing and priority rules intended for live use.

When several altcoin positions move together, portfolio drawdown can deepen even if each pair looks acceptable alone. Track open equity, peak simultaneous exposure, the number of concurrent positions and the longest capital lockup. For DCA systems, reserve room for later entries instead of pretending every pair can deploy the entire account at once.

Lock selection rules before the final test

A workable procedure has three stages. First, design the strategy and the pair eligibility rules on a development period. Second, select any settings using only that period. Third, freeze the rules and evaluate all eligible pairs on later data that has not influenced the choice. Repeatedly checking the validation period and changing the rules makes that period part of the development set.

Test sensitivity to modest changes in the liquidity cutoff, cost assumptions and review dates. Check whether a result depends on one newly listed coin or one extraordinary market phase. Forward paper trading then tests the operational path, but even clean validation does not promise future profits.

A simple reporting example

Imagine 30 pairs passed a rule written before optimization. In a purely hypothetical validation, eight produce positive net returns, 22 do not, and three winners have much higher returns than the others. Publishing only those three would hide the weak breadth. Publishing all 30 still would not tell you the portfolio return until their signals are replayed with shared capital and overlapping exposure.

The useful report includes the original eligible list, each pair's available history, settings tested, net return, maximum drawdown and trade count, plus a single account equity curve. Mark the dates on which rules and parameters were frozen. This example is educational and is not TensorParadigm performance.

• 30 pairs tested under one predefined eligibility rule.
• 8 positive and 22 nonpositive hypothetical pair results.
• Independent pair results cannot be added to estimate one account return.
• Portfolio equity and open positions require one chronological simulation.

What this means for AI trading bots

A model trained across many coins can also learn asset identity, listing era or market regime rather than a durable signal. Split data by time before fitting features and parameters, then examine results on pairs and dates that were not used to choose the model. A new coin with a short history should not receive an invented earlier track record.

Our earlier guides explain how to check trade count, detect future data leakage and measure drawdown. Pair selection adds another question: would this market have qualified before anyone knew the outcome? A clear answer makes a bot's evidence more useful than a gallery of its best charts.