Behind the Algorithms: How Stoic AI's Quant Team Builds and Tests Crypto Trading Strategies
Every Stoic AI strategy starts as a hypothesis, and most hypotheses never reach a client account. This article walks through the pipeline our quant team uses, from the first idea to live trading and, eventually, to switching a model off. It also shows where our backtested numbers end and live results begin, because that line matters more than any single return figure.
Short answer: a quant crypto trading strategy is built in six stages. A testable hypothesis, clean data, a backtest with realistic costs, validation on data the model has never seen, portfolio construction under hard risk limits, and live monitoring with predefined rules for retiring a model. Stoic AI's quant team has run this process since 2017, with more than $10 million invested in research and thousands of algorithms tested along the way.
What a crypto quant team actually does
A quantitative analyst, or quant, turns market behaviour into rules a computer can execute. The toolkit is statistics, machine learning, programming and a working knowledge of how exchanges match orders. The job is less about predicting where Bitcoin goes next and more about finding small, repeatable edges, then proving they are real and not a product of luck.
At Stoic AI the quant researchers work alongside data scientists, risk and engineering. Research decides what trades, engineering decides how it reaches the exchange, and risk decides how much any single idea is allowed to matter. If you want the building blocks of one trading algorithm (signals, entry and exit rules, position sizing), we cover them in our breakdown of a crypto trading algorithm. This article is about the process around those algorithms.
The six stages from idea to live strategy
Each stage exists to kill bad ideas cheaply, before they cost anyone money. The table shows what happens at each step and the most common reason an idea gets dropped there.
Most candidate sub-strategies are dropped at stages 3 and 4.
| Stage | What happens | What usually ends the idea |
|---|---|---|
| 1. Hypothesis | A testable claim about a market inefficiency, for example that momentum persists in liquid altcoins for several days | No economic reason why the edge should exist |
| 2. Data | Prices, volumes and funding rates are cleaned, including coins that were later delisted | History too short, or biased toward survivors |
| 3. Backtest | The rule is simulated with trading fees, slippage and funding costs | The edge disappears once costs are included |
| 4. Out-of-sample test | The rule runs on data it was never fitted to, then through historical stress periods | Results collapse outside the training window |
| 5. Portfolio and risk | The sub-strategy is sized next to everything else under hard position limits | Too correlated with what already runs |
| 6. Live and monitoring | It trades real capital and is tracked on rolling performance metrics | Its edge decays past a predefined threshold |
Stage 1: a hypothesis with a reason behind it
Ideas come from statistical and machine learning analysis of market microstructure and of how assets behave relative to each other. They fall into a few families: momentum, reversal, arbitrage, hedging and carry. The test we apply first is simple. Can we explain who is on the other side of the trade and why they keep losing to it? A pattern with no story behind it is usually a pattern in noise.
Stage 2: data that includes the losers
Crypto data is messy. Coins get delisted, exchanges go down during the moves that matter most, thin order books print wicks nobody could have traded, and funding rates change their rules. A backtest that only includes coins still listed today quietly assumes you knew in advance which ones would survive. So delisted assets stay in the data, and the tradeable universe is filtered for liquidity. Meta, for example, trades only within the top 15% most liquid assets on Binance futures and excludes memecoins and GameFi tokens.
Stage 3: a backtest that pays real costs
A backtest is a historical simulation, and the most common way to fool yourself with one is to forget costs. Strategies that trade often live or die by fees: Meta turns over around 55% of its capital per day on average, so a few basis points per trade add up fast. Every Stoic backtest includes trading fees, estimated slippage and funding. The same logic is why live results differ slightly between exchanges, which we explain in why Stoic performance can differ across exchanges.
Stage 4: out-of-sample validation and stress tests
Test enough variations of an idea and one of them will look brilliant by pure chance. This is backtest overfitting, and it is the main reason strategies that look great on paper fail live. Research by Bailey, Borwein, López de Prado and Zhu showed that standard hold-out checks are unreliable for investment backtests and proposed a way to estimate the probability that a backtest is overfit.
Our rule is that every sub-strategy passes both in-sample and out-of-sample validation before it can trade. We prefer simple rules that hold up across periods over complex ones that fit one period perfectly. Candidates are also run through historical shocks, including March 2020, May 2021, the FTX collapse in 2022 and the tariff-driven macro stress of 2025.
"Most of what we research never trades. A backtest is a hypothesis, not a result. The real question is whether the edge survives data it has never seen, honest fees and a live order book. And when a live strategy stops earning its place, we switch it off by a rule we wrote in advance, not by a debate after the fact."
Alexander Vladimirov, Head of Quantitative Research, Stoic AI
Stage 5: portfolio construction under hard limits
One good model is not a strategy. Single models break, so the products clients use are portfolios of many sub-strategies, combined so that no single idea, asset or cluster can dominate the result. Capital is allocated with a risk-based optimisation that uses historical volatility, covariance between sub-strategies and their expected Sharpe ratios.
| Strategy | Sub-strategies | Rebalancing | Position limits |
|---|---|---|---|
| Meta (market-neutral) | 200+ across momentum, reversal, arbitrage and hedging clusters; around 20 carry weight at once | Hourly | 3% per asset, 20% for Bitcoin, 40% per cluster, 40% exchange-level stop per asset |
| Stoic AI Crypto Index (long-only) | Pool of 100+; top 20 selected weekly | Sub-strategies weekly, their weights daily, asset weights hourly | 75% maximum in Bitcoin, 30% in any other single asset |
Meta keeps its long and short sides balanced dollar for dollar, so net market exposure typically stays below 1 to 2%. Its limits together cap the loss contribution of any single asset at roughly 1% of portfolio value. Automated Bitcoin Yield runs the same engine with BTC as collateral. Fixed Income is built differently: it holds offsetting spot and futures positions to collect funding payments while staying neutral to price direction.
Stage 6: live trading, monitoring and retirement
Going live is not the end of research. Every live sub-strategy is tracked on rolling performance metrics, so alpha decay, the slow fading of an edge as markets adapt, shows up early. Predefined thresholds deactivate a model whose edge has gone, and new sub-strategies join the pool when their out-of-sample results support it.
Risk is monitored around the clock, with automated alerts when a limit is approached. An Investment Committee made up of the two Co-CEOs, the CFO, the CTO, the Head of Quantitative Research and the Head of Engineering reviews performance and risk every week. Emergency actions, such as suspending a strategy or cutting exposure, are authorised by the Head of Quantitative Research and the Head of Engineering.
Backtest vs live trading: how we label our track record
Backtests are useful for research, but they are not evidence that an edge works with real money. That is why every Stoic AI strategy page states its live date, and why we split the record below instead of showing one smooth line.
| Strategy | Live since | Earlier period shown | What the earlier period is |
|---|---|---|---|
| Meta (USDT and BTC) | April 2023 | Jan 2021 to Mar 2023 | Backtest |
| Automated Bitcoin Yield | April 2023 | Jan 2021 to Mar 2023 | Backtest (same engine as Meta BTC) |
| Fixed Income | September 2022 | Jan 2021 to Aug 2022 | Backtest |
| Stoic AI Crypto Index | June 2023 (current version) | Mar 2020 to May 2023 | Live trading of the previous version |
We do not quote returns in this article because they change every month. Current figures, always shown together with Sharpe ratio and maximum drawdown, are in the monthly strategy fact sheets and on the strategy performance page.
Why a strategy gets switched off
Retiring models is a normal part of the process, not a sign that something went wrong. The usual reasons:
- Its rolling performance falls below the threshold set before it went live.
- The inefficiency got crowded, as more capital chased the same trade.
- The market changed structure, for example a long trending phase giving way to choppy, mean-reverting price action.
- Costs rose, through fee changes or thinner liquidity in the assets it trades.
- It keeps pushing against risk limits, which means it is taking more risk than its returns justify.
Risks a quant process can reduce but not remove
A disciplined process lowers the odds of trading a false edge. It does not make losses impossible. Long-only strategies fall when the market falls: Stoic AI Crypto Index has a maximum peak-to-trough decline of 68.71% over its full record, against 76.63% for Bitcoin over the same period. Market-neutral strategies carry different risks. Relationships between assets can break, both sides of a trade can lose at once, and Meta's largest decline since going live is 12%.
There is also execution risk (exchange outages, slippage in thin markets) and model risk, where a strategy keeps passing its tests until the day it doesn't. For a wider look at what rules-based trading does better and worse than discretionary trading, see our guide to algorithmic vs manual trading in crypto.
How to evaluate any quant crypto strategy
Whether you look at Stoic AI or any other provider, these six questions separate a research process from a marketing claim:
- Is the live period clearly separated from the backtest?
- Do the results include trading fees, slippage and funding costs?
- Is maximum drawdown shown next to every return figure?
- Is risk-adjusted performance, such as the Sharpe ratio, reported?
- Is there a stated process for retiring strategies that stop working?
- Do your funds stay in your own exchange account, with API keys that cannot withdraw?
We apply the same checklist in our comparison of AI crypto trading platforms. For how the Stoic app itself works day to day, from connecting an exchange to pricing, read our guide to Stoic AI's automated trading bot.
Frequently asked questions about Stoic AI's quant process
What does a quantitative analyst do in crypto trading?
A crypto quant turns market behaviour into rules a computer can execute. The work is finding small, repeatable edges, testing them on historical and unseen data with realistic costs, sizing them inside a portfolio under risk limits, and monitoring them once they trade live.
How do you know a backtest is not overfitted?
You cannot be certain, but you can reduce the risk. Stoic AI requires every sub-strategy to pass in-sample and out-of-sample validation, prefers simple rules with an economic reason behind them, includes fees and funding in every test, stress-tests against past market shocks, and keeps monitoring the strategy after launch.
How many strategies does Stoic AI run?
Clients choose from a handful of products, but each one is built from many sub-strategies. Meta draws on more than 200 sub-strategies with around 20 carrying weight at any time. Stoic AI Crypto Index selects the top 20 from a pool of more than 100 long-only sub-strategies every week.
How often are Stoic AI strategies updated?
Continuously. Meta rebalances hourly. Crypto Index reselects sub-strategies weekly, reweights them daily and adjusts asset weights hourly. Sub-strategies whose edge decays past a predefined threshold are switched off, and new ones are added when out-of-sample results support it.
Is Stoic AI's track record backtested or live?
Both, and they are labelled separately. Meta and Automated Bitcoin Yield have traded live since April 2023, Fixed Income since September 2022, and the current Stoic AI Crypto Index since June 2023. Earlier periods for Meta and Fixed Income are backtests. The earlier Crypto Index period is live trading of the previous version of the strategy.
Who oversees risk at Stoic AI?
An Investment Committee of the two Co-CEOs, the CFO, the CTO, the Head of Quantitative Research and the Head of Engineering meets weekly. Risk is monitored around the clock, and emergency actions such as suspending a strategy are authorised by the Head of Quantitative Research and the Head of Engineering.
Want to see the process in the numbers? Each strategy's monthly fact sheet shows returns, drawdowns and risk metrics side by side. When you are ready, you can connect your exchange and choose a strategy in a few minutes. Your funds stay in your own account the whole time.