Live deployment & monitoring is the final and most unforgiving stage of the edge lifecycle: the process of moving a validated strategy from backtest/paper into real capital, then continuously measuring whether it is still working. Its core tension is that a backtest is a single, frozen, hindsight-optimized sample, while live trading is an out-of-sample stream contaminated by real frictions (slippage, partial fills, latency, fees) and a non-stationary market that other participants are actively arbitraging. The job of this stage is therefore twofold — first to deploy gradually enough that you discover the gap between backtested and realized performance before it costs you, and then to monitor with predefined statistical limits so you can tell a normal, expected drawdown apart from genuine edge decay that warrants pulling the strategy.

The deployment ramp (paper → small → full)

Practitioners almost universally recommend a graduated ramp rather than going live at full size, because the first thing live trading reveals is execution reality, not strategy reality:

1. Paper / simulated live — run the production code against the live data feed with no capital. This validates the plumbing (signal generation, order routing, timestamps, data integrity) more than the edge. It will not faithfully reproduce slippage or market impact, so paper P&L is an optimistic ceiling. 2. Minimum live size — deploy with capital small enough that total loss is irrelevant. The purpose is to measure realized vs. modeled transaction costs (the "implementation shortfall" — the gap between the decision price and the achieved price). Persistent cost widening here is a red flag that the backtest's fill assumptions were fantasy. 3. Scale up — increase size only after live fills track the model and the live equity curve sits inside its expected band. Scaling also stress-tests capacity: market impact grows non-linearly with size, so a strategy profitable at $10k can be unprofitable at $1M.

A widely repeated industry figure is that the large majority of backtested strategies fail in live trading; the precise number varies by source and is not rigorously measured, so treat any single percentage as folklore-grade. The mechanism, however, is well documented (see below).

How it's monitored in practice

Monitoring replaces the comforting single backtest number with an ongoing statistical question: is live performance consistent with the distribution the backtest implied? Core tools:

• Live-vs-expected equity tracking. Plot realized equity against the backtest/walk-forward equity curve with a confidence band derived from the strategy's modeled volatility. Drift outside the band is the primary divergence signal.
• Rolling Sharpe / Information Ratio with control limits. A common heuristic is to flag when a rolling (e.g. monthly or quarterly) Sharpe falls a set fraction — frequently cited as ~30–50% — below its historical mean. The exact threshold is a chosen design parameter, not a law.
• Execution-cost delta. Track realized slippage and fees vs. the modeled assumption. A persistently widening gap means either the edge has thinned (you're trading against better-informed flow) or capacity is exceeded.
• Trade-distribution drift. Watch for P&L concentrating in fewer trades/instruments, rising max-consecutive-losses, or the win rate / payoff ratio departing from backtest. Concentration often means the diversification that smoothed the backtest is gone.
• Drawdown limits / kill-switches. A predefined maximum equity decline triggers automated de-risking: reduce size, pause, or fully halt pending manual review. The design trade-off is sensitivity — limits set too tight fire on normal noise (forcing you out of a working system at the worst time); too loose, and they fail to protect against a genuinely broken one.

Drawdown vs. broken system — the central judgment

The hardest live decision is distinguishing a drawdown a working strategy should experience from a structural breakdown. The disciplined way to frame it: a backtested Sharpe ratio implies an expected distribution of drawdown depth and duration. The CFM study "You are in a drawdown — when should you start worrying?" (Bouchaud et al., 2017) shows that under a Brownian-motion model a typical drawdown's depth scales inversely with the Sharpe ratio and its duration inversely with the square of the Sharpe — so duration is acutely sensitive to a falling edge — and argues that managers and investors chronically underestimate how deep and long a normal drawdown can be for a given Sharpe. The test is therefore whether the live drawdown is still plausible under that distribution. When a drawdown exceeds what the assumed Sharpe makes credible, the correct response is to revise the assumed Sharpe downward — i.e. update your belief that the edge is intact — rather than to assume mean reversion. A common practical caveat: the clean closed-form math assumes IID/Brownian returns, but real P&L is serially correlated, so these bands are guides, not proofs (the CFM paper itself adds an autocorrelation correction).

Adoption, debate & evidence

Graduated deployment and live monitoring are standard institutional practice and are the explicit subject of regulatory/compliance documentation for algorithmic trading (paper-trading results, backtest-vs-live comparison). The deeper, evidence-backed claim is why monitoring is mandatory: edges decay, and the decay is measurable.

The strongest evidence comes from McLean & Pontiff (2016, Journal of Finance), who studied 97 published cross-sectional return predictors and found portfolio returns were 26% lower out-of-sample (after the original sample period but before publication) and 58% lower post-publication. The authors attribute the 32-percentage-point difference (58% − 26%) to publication-informed trading — investors learning about the mispricing from the academic paper itself. This separates two distinct decay sources: statistical (overfitting/data-mining), captured by the out-of-sample drop, and economic (crowding/arbitrage), captured by the further post-publication drop. Both are reasons live results systematically undershoot backtests. A genuine debate persists over how much is overfitting vs. real-but-arbitraged edge — McLean & Pontiff explicitly frame the 26% out-of-sample decline as an upper bound on data-mining effects.

Strengths & limitations

Disciplined deployment-and-monitoring is the only mechanism that catches the two failure modes a backtest cannot: overfit edges that were never real, and real edges that the market has since competed away. It works because it forces predefined exit criteria, removing the in-the-moment rationalization that keeps traders in dead strategies.

Its limits are statistical. Equity-curve significance is slow — distinguishing a 30% Sharpe drop from noise can require many trades/months, by which time substantial capital may be lost; conversely, tight limits cause whipsaw, ejecting you from a sound strategy mid-drawdown. The single most common misuse is monitoring without pre-committed thresholds — judging the live curve by feel, which collapses into hindsight bias in both directions (over-trusting a lucky run, abandoning an unlucky-but-valid one). A second frequent error is treating paper-trading P&L as predictive of live P&L despite its missing slippage and impact.

System relevance

Within Delvantic this node is the conceptual bridge between the Augustus trade-setup agent (which proposes setups) and Cairn (which holds the measured live track record). Augustus should treat any historical/backtested win rate as an upper-bound prior, not a live expectation, and defer to Cairn's realized statistics where they exist — exactly the out-of-sample haircut McLean & Pontiff quantify. The hard caveat for the agent: backtest confidence is not live confidence; live confidence is earned only through the monitored, out-of-sample record.

Sources

• McLean, R. David & Pontiff, Jeffrey (2016), "Does Academic Research Destroy Stock Return Predictability?", Journal of Finance 71(1), 5–32 — 26% out-of-sample / 58% post-publication decline. (papers.ssrn.com/abstract=2156623; Wiley onlinelibrary.wiley.com/doi/abs/10.1111/jofi.12365)
• Bouchaud, J.-P. et al. / Capital Fund Management (2017), "You are in a drawdown. When should you start worrying?" — primary source for the Sharpe-implied drawdown distribution: under Brownian motion, drawdown depth ∝ 1/Sharpe and duration ∝ 1/Sharpe²; managers underestimate normal drawdown depth/length; includes an autocorrelation correction. (arxiv.org/pdf/1707.01457; cfm.com)
• Harbourfront Technologies, "When Trading Systems Break Down: Causes of Decay and Stop Criteria" — overfitting vs. arbitrage decay; drawdown-vs-breakdown framing (secondary, popularizing the CFM result). (derivvaluation.medium.com)
• Surmount.ai, "Walk-Forward Analysis vs. Backtesting" — graduated deployment, monitoring live vs. walk-forward predictions.
• PineConnector, "Backtesting vs Live Trading: Bridging the Gap" — backtest-to-live gap, slippage/fills.
• QuantVPS, "How to Create a Trading Algorithm: Design, Backtest & Deploy" — paper → scaled deployment ramp, dashboards/alerts.
• Tradetron / Nurp, algo risk-management guides — drawdown control limits, kill-switches, rolling-Sharpe control-limit heuristics.

Disputed/qualified: the "majority of backtests fail live" figure is widely cited but not rigorously measured — treated as folklore. Specific Sharpe-drop and drawdown thresholds are design choices, not established constants. Sharpe-implied drawdown bands assume near-IID returns, which real P&L violates.