Getting a language model to place a trade is now the easy part. Mint a key, wire an MCP server or a ChatGPT Custom Action, and within ten minutes an agent can read a price and call an order endpoint. The part nobody hands you a walkthrough for is the sixty seconds before that call — the actual rule that decided this trade, now, this size was worth making. That is strategy, and it is the one piece of a trading agent you cannot outsource to a runner, a scope, or a persona template.
This article is the strategy layer, not the build layer. If you have not wired an agent yet, Build Your Own Crypto Trading Agent covers the MCP and API-key walkthrough end to end, and designing your agent covers the perceive-decide-act loop and the operating principles a disciplined agent follows no matter which strategy it runs. If you have not read the category explainer, start with What Is Agentic Trading?. What follows here is six crypto trading bot strategies archetypes an agent can actually implement: trend and momentum following, mean reversion, news and event-driven trading, cross-market signals, calendar and catalyst strategies, and portfolio rebalancing. Each gets its ai trading agent strategy hypothesis, the data it needs, how to express it as agent trading rules, and — deliberately — its honest failure mode. This is trading agent strategy design with the losing side included on purpose: no llm trading strategies archetype below is presented as a money-printer, because every one of them breaks in a specific, predictable way, and knowing that failure mode in advance is most of what separates a real strategy from a fantasy.
Ground truth before you read further: everything below is scoped to paper trading strategies only. A CoinRithm paper account trades a virtual 50,000 mUSD balance against live market prices — never real capital. Nothing here is a backtest result, a performance guarantee, or financial advice. Several sections exist specifically to describe how and when a plausible-sounding strategy loses money, on paper, so you see it before you trust it with anything more.
TL;DR
- Six strategy archetypes an agent can run and honestly evaluate: momentum, mean reversion, news/event-driven, cross-market signals, calendar/catalyst, and portfolio rebalancing.
- Each one gets a hypothesis, its data needs, how to write it as agent rules, and its known failure mode — the failure mode is the point, not an afterthought.
- Strategy, execution, and risk are three separate jobs. A good strategy with no risk caps still eventually blows up on one bad regime call.
- A fair paper-trading test needs enough decided trades, out-of-sample time, and a start date you didn't pick after already seeing how it played out.
- CoinRithm evaluates agents on realized PnL and decided trades via the public Agent Arena — open positions don't count until they close.
- Paper only, always. Nothing here predicts real-money performance, and this is not financial advice.
Trend and Momentum Following
The hypothesis: an asset that has been moving in one direction keeps moving in that direction longer than pure chance would predict — participation lags information, and a real move tends to attract more of it before it exhausts. A momentum strategy ai agent is betting on continuation, not reversal.
- Data it needs: OHLCV candles across a couple of timeframes, 24h/7d price change, a rolling volume trend, and a news feed ranked by importance and recency.
- As agent rules: something like
if price is above its 20-period high AND 24h volume is rising AND no high-importance bearish news in the last 6h, treat this as a trend regime and consider a long. The part an LLM adds that a fixed indicator can't: regime classification from news context — deciding whether a breakout is a genuine shift (a real catalyst, confirmed) or noise (a thin-volume spike, an unconfirmed rumor) before treating it as trend rather than treating every crossed threshold identically. - Known failure mode: chop. In a range-bound market, a momentum rule fires on every minor swing, gets stopped out on the reversal, and fires again on the next swing the other way — a string of small, fee-bearing losses that no single trade explains but that add up fast. The regime classification step exists specifically to reduce this, and it will still be wrong sometimes.
- A fair paper-trading test: enough decided trades to include both a trending stretch and a choppy one — not a backtest window quietly chosen because it was mostly trending. A momentum strategy tested only during one strong bull run will look far better than it is.
Mean Reversion
The hypothesis: price overshoots relative to its recent average and tends to snap back toward it — the same overshoot-then-correct pattern that shows up across liquid markets when nothing structural has changed. A mean reversion agent buys dips and sells rips inside a range.
- Data it needs: a rolling mean and standard deviation band, the recent trading range (high/low), and — critically — the same news feed the momentum strategy uses, for a different reason: to rule out a real catalyst before assuming "this is just noise."
- As agent rules: something like
if price is more than 2 standard deviations below its 20-period mean AND no news catalyst explains the move, consider a reversion long toward the mean; exit at the mean or at a hard stop. - Known failure mode — and this is the one that matters most: mean reversion blows up in a real trend, specifically. There is no mean to revert to yet when a market is repricing to genuinely new information — the "cheap" dip an agent buys is not cheap, it's early, and a rule that keeps adding to a reversion position as price keeps falling is the textbook way a small, sensible-looking bet becomes a large, ugly one. The fix isn't a smarter reversion rule; it's the same regime question from the section above, asked in reverse — a mean-reversion agent needs a trend filter that tells it when not to fire at least as much as it needs the reversion signal itself.
- A fair paper-trading test: the same discipline as momentum, inverted — enough decided trades across both a genuine range and a genuine trend, so the strategy's behavior during the trend (does it stop firing, or does it keep buying into it) is actually visible in the record instead of hidden by a lucky test window.
News and Event-Driven Trading
The hypothesis: markets don't instantly and perfectly price new information — there is a short window after a headline where an agent that reads and reasons about it faster or more carefully than the crowd can act before the price fully catches up.
- Data it needs: a watchlist news feed ranked by importance with a real publish timestamp (not a cache timestamp), and current price context to check whether the market has already moved.
- As agent rules: something like
if a high-importance item published in the last 30 minutes hasn't yet been matched by a proportional price move, consider a small position sized down for uncertainty; state the specific headline and the specific reason, not "sentiment is bullish". - Known failure mode: this is the archetype where an LLM's biggest natural edge — reading unstructured text — is also its biggest trap. A model can hallucinate a plausible-sounding but wrong detail, conflate an unconfirmed rumor with a confirmed fact, or act on stale context: a headline it's reasoning about that the market already priced in an hour ago, because the feed it read was cached or because it simply took too long to decide. Acting confidently on stale or invented context is a worse failure than doing nothing, because the trade looks reasoned right up until it isn't.
- A fair paper-trading test: check the timestamp gap between the news item and the trade decision on every entry in the log, not just the win/loss outcome — a strategy that "works" but is actually trading three-hour-old news some of the time isn't a repeatable edge, it's a data-freshness bug that happened to pay off.
Cross-Market Signals: Prediction Markets Meet Crypto
The hypothesis: a prediction market's probability on a relevant event — a macro decision, a listing, a regulatory outcome, a price-threshold market like "will BTC hit $150k by year-end" — can function as an input to crypto positioning, and the reverse also holds: a sharp crypto price move can be a leading signal for a related event market that hasn't caught up yet.
- Data it needs: crypto market context alongside related prediction-market odds for the same underlying event — CoinRithm's market-context data already links the two, so an agent doesn't need to hand-build the mapping.
- As agent rules: something like
if a linked prediction-market probability shifts by more than 10 points in a day and the matching crypto price hasn't moved proportionally, treat that as an input to conviction — not a stand-alone trigger on its own. Treat it as one signal among several, exactly the discipline covered in How AI Agents Trade Prediction Markets, applied here in the other direction. - Known failure mode: a
prediction market agent strategybuilt on a cross-market gap can mistake noise for signal twice over — a thin prediction market can be moved by a single trade, so the "shift" an agent reacts to may be one participant's opinion, not a crowd's; and two markets that look like they're pricing the same question can resolve on subtly different wording, so an apparent mispricing between them is often two correctly-priced but different bets, not one exploitable gap. - A fair paper-trading test: never score this archetype as a stand-alone bet-for-bet strategy on its own trade log; score whether including the cross-market input changed the quality of the crypto-side decisions across many cycles, since the entire hypothesis is that it's a conviction input, not an independent signal.
Calendar and Catalyst Strategies
The hypothesis: some volatility is foreseeable in advance — a scheduled macro release, an options-expiry date, a token-unlock schedule, a prediction market's own resolution deadline — and an agent that knows the calendar can plan around the date instead of being surprised by it.
- Data it needs: an events calendar (macro dates, on-chain unlock schedules, prediction-market close and resolution dates) cross-referenced against current positions.
- As agent rules: something like
if a scheduled high-impact catalyst falls within the next 24 hours, reduce new position size by half, or wait until after the event for directional clarity before sizing normally again. - Known failure mode: "sell the news." A well-telegraphed event is often already priced in by the time it arrives, so a naive calendar rule that assumes the catalyst causes a directional move can get the direction backwards precisely when the outcome matches consensus. Calendar data can also slip — a vote reschedules, a decision is delayed — so a rule keyed to a fixed date needs to treat a stale calendar entry as a reason to abstain, not act.
- A fair paper-trading test: requires multiple instances of the same type of catalyst (several FOMC dates, several unlocks), not one lucky or unlucky pass through a single event — one data point can't tell you whether the rule works or whether you just watched one outcome.
Portfolio and Rebalancing Logic
The hypothesis: holding a basket at fixed target weights — across coins, or across venues like spot, futures, and prediction markets — and periodically trading back to those weights captures a systematic "trim the winner, top up the laggard" effect without needing any directional view at all. This is an ai agent portfolio strategy, not a single-asset bet.
- Data it needs: current position weights across the whole paper portfolio, a target allocation, and a rebalancing band (how far a weight can drift before it triggers a trade).
- As agent rules: something like
every cycle, compute each asset's share of total portfolio value; if any asset drifts more than 5 percentage points from its target weight, submit a resizing trade back toward target. - Known failure mode: the same trend-versus-mean-reversion tension from earlier shows up again here, at the whole-portfolio level. Rebalancing is a mean-reversion bet on relative weights — it sells the position that's been running and buys the one that's been lagging — so in a market where one asset trends hard for a long stretch, disciplined rebalancing quietly bleeds returns by repeatedly trimming the winner. Tight rebalancing bands with frequent trades compound this with real, disclosed execution costs (spread, slippage, taker fees) on every resize.
- A fair paper-trading test: always compare against a static buy-and-hold baseline over the identical window, and account for every fill's disclosed cost — a rebalancing strategy that "wins" only when compared to nothing is not actually evidence of anything.
Strategy, Execution, and Risk Are Three Different Jobs
Everything above is the strategy layer: the hypothesis, the data, and the rule that turns a read into a decision. It is not the whole agent, and treating it as the whole agent is how a reasonable-sounding idea does real (paper) damage.
- Strategy decides what and when — the six archetypes above.
- Execution decides how the decision actually becomes an order: resolving a symbol to the right ID, quoting before every trade, sizing the position, and using idempotency keys so a retry never double-fills. This is the layer Build Your Own Crypto Trading Agent walks through in detail.
- Risk decides what the strategy is not allowed to do regardless of how confident it sounds: a maximum position size, a maximum leverage, a daily loss limit, a cap on open positions, a kill switch on repeated failures.
A good strategy with no risk caps still eventually dies — not because the hypothesis was wrong, but because every one of the failure modes above will happen eventually, and without a hard cap in place, one bad regime call, one hallucinated headline, or one thin cross-market read can size far larger than it should. Strategy quality and risk discipline are independent variables; a mediocre strategy with tight caps usually survives longer than a clever one without them. For the caps, sizing math, drawdown limits, and kill-switch mechanics that make any of the six archetypes above survivable, see Trading Agent Risk Management.
Evaluating a Strategy Honestly on CoinRithm
A strategy hasn't been tested until it's been tested honestly, and honesty here has a specific shape:
- Decided trades, not open positions. A position that hasn't closed hasn't proven anything yet — it could still go either way. Judge a strategy on realized outcomes, the same rule the Agent Arena enforces publicly.
- Enough decided trades to matter. Five trades is an anecdote. A strategy needs a large enough sample of closed positions before a win rate or an average return means anything statistically.
- Out-of-sample time. If you tuned the prompt while watching how it performed, that period is not a test — it's the training data. The honest test is the stretch of time after the rules were frozen.
- No cherry-picked start dates. Pick the evaluation window before you know how the strategy performed in it, not after. A mean-reversion agent tested only during a calm, ranging month will look far better than it is; the same is true in reverse for momentum tested only during one strong trend.
- A public track record. Anyone can describe a strategy that "would have worked." A strategy is proven, not described, when its decided trades are checkable — which is exactly what the Agent Arena's realized-PnL ranking is for. If you'd rather test the same discipline yourself before automating any of it, CoinRithm's own paper trading product runs the identical mUSD sandbox without needing an agent at all.
How CoinRithm Fits In
Every archetype above maps directly onto surfaces CoinRithm already runs as a paper-trading environment, not a hypothetical one:
- Market context and news, in one read. The data each strategy needs — price, candles, 24h/7d change, ranked news, and linked prediction-market context — comes back from a single market-context call, so an agent isn't stitching together five separate sources to run one rule.
- A keyed agent API and MCP server for the execution layer. Once a strategy has decided, the same
crk_live_key and MCP tool set covered in Build Your Own Crypto Trading Agent handles the resolve-quote-trade mechanics for spot, futures, and prediction-market positions from one paper balance. - The Agent Arena as the honesty mechanism. Realized PnL, decided trades, and rolling time windows — the same evaluation discipline described above, made public and comparable across every opted-in agent.
- Paper only, end to end. Every archetype here trades a virtual mUSD balance against live prices. No real money, wallet, or exchange account is involved at any point in any of the six strategies described.
None of this is a claim that any strategy above is profitable, on CoinRithm or anywhere else. It's a description of how to build one, how it fails, and how to find out honestly which is true for the version you built.
Frequently Asked Questions (FAQ)
What's the difference between a trading strategy and a trading agent's execution logic?
Strategy decides what to do and when — the hypothesis, the data it reads, and the rule that turns a read into a decision. Execution is how that decision becomes an actual order: resolving the right symbol, quoting before the trade, sizing it, and handling the fill safely. A good strategy can still lose money through bad execution, and solid execution can't rescue a bad strategy — they're separate jobs that both have to work.
Can any of these six strategies guarantee profit?
No. Every archetype above has a known, honest failure mode described in its own section — that's deliberate, not a hedge. Momentum chops in ranging markets, mean reversion compounds losses in real trends, news trading can act on stale or hallucinated context, cross-market signals can mistake thin-liquidity noise for a real gap, calendar strategies can get "sell the news" backwards, and rebalancing bleeds returns in a strong trend. Nothing here is financial advice, and no result — paper or otherwise — predicts future performance.
Which strategy archetype should a beginner agent start with?
Whichever one you can state as a specific, falsifiable rule rather than a vibe — "buy when X and Y are both true" beats "buy when it looks strong." Momentum and portfolio rebalancing tend to have the simplest data requirements to start with. Whatever you pick, set risk caps before you set the strategy live; a good beginner strategy with hard caps beats a clever one without them.
How many decided trades count as a fair test of a strategy?
More than a handful, and specifically enough to cover more than one market regime — a trending stretch and a ranging one, at minimum. A strategy tested only during the regime it's suited for (momentum during a bull run, mean reversion during a calm range) will look better than it actually is. Judge it on realized, closed trades, not open positions, and be honest about whether the evaluation window was picked before or after you saw the result.
Why does mean reversion specifically blow up in trends?
Because mean reversion is a bet that price returns to a recent average, and in a genuine trend there is no average to return to yet — the market is repricing to something new. A reversion rule that keeps buying "cheap" dips as a real downtrend continues is adding to a loser precisely because its own signal (further below the mean) keeps firing. The fix is a trend filter that tells the strategy when to stay out, not a better reversion signal.
Can an agent combine multiple strategy archetypes at once?
Yes, and in practice most real agents do — for example, using calendar awareness to size down before a scheduled catalyst while otherwise running a momentum rule, or using a cross-market signal as one input alongside price action rather than as its own trigger. Combining archetypes doesn't remove any of their individual failure modes, though — it just means the risk layer has to account for more than one kind of wrong at once.
Continue reading: Trading Agent Risk Management — the caps, sizing math, drawdown limits, and kill-switch mechanics that keep any strategy above from turning one bad call into an account-ending one.
Disclaimer: This article is for educational purposes only and is not financial or investment advice. All trading described here uses simulated mock USD (mUSD) on CoinRithm; no real money, wallet, or exchange account is involved at any step. Nothing in this article predicts or guarantees any strategy's performance, and paper-trading results — for any of the six archetypes described — do not predict real-trading performance.
