Signal Decay in Trading: When a Backtested Edge Vanishes Before the Fill

Signal Decay in Trading: When a Backtested Edge Vanishes Before the Fill

বাংলায় পড়ুন

Evergreen trading tutorial: This guide uses hypothetical data to explain how timing and execution can reduce a trading signal's apparent advantage. It is educational information, not personalised advice or a promise of returns.

A backtest may detect an attractive signal at 16:00, value the entry at that same closing price and report a profitable result. A real trader sees the completed data later, sends an order, waits in a queue and fills at a different price. By then, part or all of the expected advantage may have disappeared.

This loss of predictive value through time is signal decay. It connects research, market structure and trading costs. A strategy is usable only if enough of its gross edge survives the journey from observation to executable fill.

Learning objective

By the end of this guide, you should be able to:

• distinguish signal time, decision time, order time and fill time;
• measure how expected return changes under realistic delays;
• estimate a signal's practical half-life without treating it as permanent;
• combine decay with spread, slippage and fees; and
• reject or redesign a strategy whose edge cannot survive execution.

The four timestamps every test needs

1. Signal time

The latest moment when all inputs are genuinely known. A rule using the official daily close cannot know that close before the closing auction is complete.

2. Decision time

The moment the system finishes calculations and decides what to trade. Data delivery, cleaning and model computation create delay.

3. Order time

The moment the broker or exchange receives the instruction. Network, software and human approval add more delay.

4. Fill time

The moment the order actually executes. A limit order may wait or never fill; a market order may fill immediately but at a worse price.

If a backtest uses a price from before the signal was knowable, it contains look-ahead bias. If it uses a theoretically available price that real orders rarely obtain, it has an execution assumption problem.

What signal decay means

Suppose a signal predicts a positive return over the next hour. Other traders react, liquidity providers adjust quotes and new information arrives. The forecast can weaken with every minute.

Fast decay is common when the idea depends on short-lived order flow, a news reaction or a temporary price dislocation. Slower decay may appear in multi-day trend, valuation or fundamental signals. Speed alone does not make one strategy better: faster systems usually face greater spread, slippage, infrastructure and market-impact demands.

The useful quantity is net executable edge, not the backtest's gross return.

Net edge = gross expected return at the achievable fill time − spread − slippage − fees − financing and other costs

Signal half-life

A signal's half-life is the delay at which its estimated expected return falls to half the value measured at the earliest valid entry. If an earliest-valid-entry estimate is 0.20% and the estimate falls to 0.10% after 15 minutes, the observed half-life is about 15 minutes.

This is a sample estimate, not a physical constant. Half-life can change with volatility, liquidity, competition, asset, session and regime. A smooth exponential-decay curve may be convenient but should not be assumed unless the data support it.

Worked hypothetical example

Assume a researcher tests 200 historical signals using only data known at the end of each five-minute bar. The average gross return to a fixed exit changes with entry delay:

Entry assumption	Average gross return
First executable quote	+0.24%
5-minute delay	+0.16%
15-minute delay	+0.10%
30-minute delay	+0.04%

Estimated round-trip spread, slippage and fees total 0.07%. Approximate net returns become:

• first executable quote: +0.17%;
• 5-minute delay: +0.09%;
• 15-minute delay: +0.03%; and
• 30-minute delay: −0.03%.

The research signal remains directionally positive after 30 minutes, but the tradable strategy is negative after estimated costs. If the real operating delay is normally 15 to 25 minutes, the earliest result is not a realistic performance estimate.

The fall from 0.24% to near 0.12% occurs between 5 and 15 minutes, suggesting a rough half-life in that interval. A confidence interval and unseen-data test are still required.

Step-by-step delay test

1. Timestamp every input

Record when each price, indicator, filing or economic release became available. Use publication time rather than the period the data describe.

2. Define the earliest valid entry

Allow enough time for data delivery and calculation. For a close-based signal, the same close may be impossible unless the rule can be submitted before the auction using information that excludes the final close.

3. Re-run with an execution ladder

Test several realistic delays: next quote, one bar, five minutes, 15 minutes, next open or next session, depending on the strategy. Keep the exit rule unchanged so the delay effect is visible.

4. Use executable prices

For buys, consider the ask and for sells, the bid. Model market-order slippage or limit-order fill probability. Mid-price fills are often optimistic.

5. Add total costs

Include commission, spread, slippage, exchange fees, currency conversion, borrow cost and financing where relevant. Our bid-ask spread guide covers the main execution frictions.

6. Segment before inspecting

Predefine liquidity, volatility, market session and instrument groups. A signal may survive in liquid large-caps but vanish in thin shares.

7. Validate later data

Choose any speed requirement on a research sample, then keep it fixed on unseen trades. Re-optimising the delay threshold on the test set creates another form of overfitting.

8. Compare paper and live fills

Record decision price, arrival price, fill price and rejected or missed orders. Our implementation-shortfall guide explains how to attribute this difference.

Limit orders do not remove decay

A limit order controls the worst acceptable price but does not guarantee execution. If price moves favourably without filling, the backtest must not assume the trade happened. If the order fills mainly when price is about to move against it, the strategy experiences adverse selection.

A realistic limit-order test needs rules for queue position, partial fills, cancellation and what happens after a miss. Counting only completed favourable fills overstates performance.

Capacity and crowding

A signal can decay faster when more capital trades it. Larger orders take longer to complete, consume more order-book depth and reveal demand. A strategy that works with £10,000 may not scale to £10 million at the same return.

Capacity should therefore be tested with participation limits and market impact. Read our strategy-capacity guide for the size dimension.

Common mistakes

1. Entering at the same close used to calculate the signal. The price may precede the completed information.
2. Using bar highs or lows as guaranteed fills. A touched price does not prove the order's queue position executed.
3. Measuring decay before costs only. A positive gross return can already be negative net.
4. Assuming zero operational delay. Data, decisions and orders each take time.
5. Ignoring missed limit orders. Non-fills are part of the strategy outcome.
6. Optimising one perfect delay. Many tested delays create selection bias.
7. Combining liquid and illiquid assets. The average can hide where execution is impossible.
8. Treating half-life as permanent. Competition and regimes change it.
9. Using current infrastructure in an old backtest without adjustment. Historical data availability may differ.
10. Scaling capital without retesting impact. More size can extend execution beyond the signal's life.

Practical checklist

• Every input has a genuine availability timestamp.
• Signal, decision, order and fill times are stored separately.
• The earliest backtest entry is operationally achievable.
• Several realistic delays are tested.
• Bid, ask, spread, slippage and fees are included.
• Limit-order misses and partial fills are counted.
• Delay tests are segmented by liquidity and session using pre-set rules.
• Half-life estimates include uncertainty and sample size.
• Delay assumptions are validated out of sample.
• Live arrival and fill prices are compared with research assumptions.
• Capacity and market impact are retested as size changes.
• The strategy remains acceptable after a conservative execution scenario.

The Money Man View

A signal is not a trade until an executable order fills. Research should begin at the earliest valid timestamp, then deliberately add the delays and costs the live process will face. If the result collapses after one realistic bar or a modest spread, the apparent edge is too fragile. Build around the delay you can reliably achieve and demand a margin of safety rather than engineering a backtest around an unavailable price.

Educational risk disclosure: Backtests and delay estimates cannot guarantee future execution or returns. Liquidity, costs and signal behaviour can change abruptly, and orders may not fill. The figures above are simplified hypothetical examples. This article was prepared with AI-assisted research and writing under The Money Man editorial process.