Risk management in crypto trading is almost universally misunderstood. Most traders equate it with setting a stop-loss. A stop-loss is one mechanism that addresses one problem: limiting loss on one open position. Real risk management is a multi-layered system that addresses position sizing, portfolio correlation, strategy decay, operator absence, crash recovery, and systemic stress.
We run 45 automated bots managing 45,000 dollars across 6 strategy types. Our risk framework has eight distinct layers, each addressing a different failure mode. Here is the complete architecture.
Layer 1: Stop-Loss Enforcement
Every trade must have a predefined stop-loss before entry. This is enforced in code: if a strategy generates a signal without a suggested stop-loss price, the order is rejected. Stop-losses are set using ATR multiples (typically 1.5 to 2.0 times ATR) which automatically adjust to current volatility. During calm markets, stops are tighter. During volatile markets, stops are wider.
ATR-based stops are superior to fixed-percentage stops because they account for the asset's current volatility. A 2 percent stop on BTC makes sense when daily range is 1.5 percent but is too tight when daily range is 8 percent. ATR normalization handles this automatically.
Layer 2: Position Sizing
Maximum position size is 25 percent of each bot's allocated capital. This prevents any single trade from being large enough to materially damage the bot's equity in a single loss.
Within this cap, half-Kelly sizing determines the actual position. The Kelly fraction (calculated from the strategy's win rate and payoff ratio) is multiplied by 0.5 for conservative sizing. This balances growth rate (roughly 75 percent of theoretical maximum) against drawdown risk (roughly half the variance of full Kelly).
The correlation sizer further adjusts based on portfolio overlap, scaling positions between 25 percent and 100 percent of their intended size depending on how correlated the new position is with existing portfolio exposure.
Layer 3: Drawdown Circuit Breaker
Each bot has a 20 percent maximum drawdown from peak equity. Hourly equity snapshots track the high water mark and current equity. When the decline from peak exceeds 20 percent, all new entries are blocked for that bot. Existing positions can still be closed but no new risk is added.
A 20 percent drawdown requires a 25 percent gain to recover. This is achievable for our strategies (validated Sharpe 1.7 to 19.0) but represents significant capital impairment. The 20 percent threshold was calibrated using Monte Carlo analysis: it exceeds the 95th percentile drawdown for our deployed strategies while providing a buffer for unusual market conditions.
Layer 4: Daily Loss Limit
If a bot's realized losses for the current UTC day exceed 5 percent of its allocated capital, all trading suspends until the next day. This prevents catastrophic intraday loss sequences where multiple stop-losses fire during a crash.
The 5 percent daily limit is independent of the drawdown breaker. A bot can lose 5 percent on day one (daily limit triggers), 5 percent on day two (triggers again), and reach the 20 percent drawdown breaker on day four. Each mechanism catches a different loss pattern: the daily limit catches acute intraday losses, the drawdown breaker catches sustained degradation.
Layer 5: Consecutive Loss Cooldown
After three consecutive losing trades, position sizing halves for each additional loss. The multiplier after 3 losses is 0.50, after 4 losses is 0.25, after 5 losses is 0.125, with a floor at 10 percent. This progressive reduction means a strategy in a hostile regime automatically throttles itself rather than continuing to lose at full size.
The multiplier resets after a single winning trade. This mechanism is specifically designed for regime transitions: when a strategy enters conditions where its edge does not apply, the losses come in sequences, and the cooldown reduces exposure until either conditions improve (a win resets the multiplier) or the decay detector pauses the bot entirely.
Layer 6: Portfolio Constraints
Individual bot risk management is necessary but not sufficient. Three portfolio-level constraints prevent correlated failures across all 45 bots.
Total exposure cap: 50 percent of aggregate capital (22,500 dollars in open positions maximum). Single asset concentration: 25 percent of aggregate capital (no more than 11,250 dollars on any single symbol). Portfolio drawdown halt: 15 percent aggregate decline from peak (all trading stops, manual reset required).
These constraints prevent the scenario where 30 individually risk-managed bots simultaneously go long on correlated altcoins during a euphoria phase, creating catastrophic portfolio-level exposure.
Layer 7: Strategy Decay Detection
A strategy that passes validation today might stop working in three months. The decay detector monitors rolling 30-day Sharpe for each bot. With at least 10 trades in the window (minimum for statistical significance), it calculates the annualized Sharpe using per-trade PnL percentages.
When the rolling Sharpe drops below 0.5, the bot is automatically paused. A Sharpe of 0.5 means the strategy is barely outperforming zero after risk adjustment. Our deployed strategies have validated Sharpes from 1.7 to 19.0, so 0.5 represents severe degradation that warrants intervention.
The 10-trade minimum prevents false positives. A bot with only 3 recent trades does not have enough data for a meaningful assessment. The detector only activates with sufficient statistical evidence.
Layer 8: Dead Man's Switch
The only risk mechanism that does not depend on market conditions. The dead man's switch requires operator check-in via Telegram every 24 hours. The escalation sequence: warning at 83 percent of timeout (approximately 20 hours), critical at 96 percent (approximately 23 hours), full shutdown at 100 percent.
The latch mechanism ensures that after trigger, a simple check-in does not restart the system. The operator must send an explicit reset command, confirming they have reviewed the situation. This prevents automated scripts from masking genuine operator absence.
How the Layers Interact
The eight layers form a defense in depth. Any single layer can fail and the others still protect the portfolio. If the stop-loss fails (exchange lag, flash crash through the level), the daily loss limit catches the damage. If the daily loss limit is insufficient (multi-day decline), the drawdown breaker fires. If the drawdown breaker is too slow (strategy producing many small losses), the consecutive loss cooldown throttles sizing. If individual bot limits are all within range but the portfolio is dangerously concentrated, the portfolio constraints block new entries. If the strategies are slowly degrading without triggering any individual bot limit, the decay detector catches the trend. If everything else fails because no one is monitoring, the dead man's switch stops the system.
No single mechanism handles all failure modes. The layered approach ensures that each mode is addressed by at least one dedicated mechanism, and most modes are addressed by multiple mechanisms simultaneously.
The Cost of Risk Management
Risk management reduces returns. The portfolio constraints prevent maximum deployment of capital. The correlation sizer reduces positions that would otherwise be larger. The consecutive loss cooldown throttles strategies that might recover quickly. The dead man's switch stops trading during the 24-hour check-in gap.
This cost is real but acceptable. The alternative — maximum capital deployment with no constraints — produces higher returns in favorable conditions and catastrophic losses in unfavorable ones. Risk management trades expected return for survivability, which is the correct trade for any system managing capital with a long time horizon.
Our validated Sharpe ratios (1.7 to 19.0) are measured after all risk layers are applied. The numbers would be higher without constraints, but they would also be meaningless because the unconstrained system would not survive the drawdowns it encounters during adverse regimes.