← All journal articles

A Smaller Replay Loss Is Not Proof of a Profitable Strategy

TradeStack's October 16 controlled replay compared two entry-quality filters with a baseline. Here is what the reported numbers mean—and what they cannot establish.

A common mistake in systematic trading research is to find an improvement in a backtest and immediately call it an edge.

Our October 16, 2026 research experiment is a useful counterexample. An additional entry-quality filter improved the *reported* after-cost replay outcome. But the test was a partial session with only two closed trades per variant—and every variant remained net negative.

This article summarizes a private TradeStack research report, not a fresh independent rerun of the market recording.

What we tested

The report compared the frozen strategy-v2026.10.15 pipeline with two optional checks:

  • Immediate one-minute quality: at the decision time, reject a directional candidate when completed one-minute data indicates adverse EMA5/MACD-histogram slope or a two-candle exhaustion/rejection pattern.
  • Completed five-minute alignment: require the last completed five-minute candle to agree with the intended trade direction and EMA21 relationship.

Both variants used only information available at the original decision timestamp. The experiment patched the research process temporarily; it did not permanently change live or paper strategy configuration.

What the repository report recorded

The controlled comparison covered 09:15–11:40 IST on October 16, with 57,579 callback events, 380 analyzer evaluations, depth-based fill assumptions, a 100 ms modeled latency and consistent fee assumptions across variants.

VariantClosed tradesWin rateGross P&LNet P&L after costs
Existing baseline20%−₹1,240.50−₹1,459.77
Immediate one-minute quality250%−₹273.00−₹491.05
Completed five-minute alignment20%−₹1,240.50−₹1,459.77

The improvement in reported net P&L from baseline to the immediate filter is:

(-491.05)-(-1459.77)=+₹968.72

The ₹968.72 difference is not profit. It is a reduction in the magnitude of a simulated loss.

The arithmetic behind the reported gross/net differences can also be checked: baseline implied costs were ₹219.27, while the one-minute variant implied costs were ₹218.05. Absolute fees remain uncertain because the study noted a mismatch between local and broker estimates.

Why we cannot simply remove losing trades

A filtered signal changes the rest of a trading session.

The research report says the immediate filter rejected replay opportunities around 09:33, 09:58 and 10:56. Later entries at 09:46 and 11:07 then differed because session state had changed.

That state includes active positions, post-stop cooldowns, capital availability and drawdown boundaries. Deleting the rejected trades from a spreadsheet would not re-create the counterfactual session.

This is why the study used an in-process replay variant rather than subtracting historical losses from the trade ledger.

What the five-minute filter taught us

The completed five-minute alignment requirement did not improve reported after-cost results in the main window. Early losing trades still occurred.

More confirmation is not automatically better. The extra time needed to form a completed candle may leave poor entries unchanged while introducing delay to other opportunities.

The report did not recommend promoting this filter.

Limitations are part of the result

The study documented several reasons not to convert the observed improvement into a production rule:

  1. Only two partial sessions were usable for cross-session consideration; the main comparison had just two closed trades per variant.
  2. Paper and replay execution did not fully match for October 16.
  3. The replay trade setup field was labeled UNKNOWN, limiting setup-level attribution.
  4. Local transaction fees did not fully agree with broker estimates.
  5. An October 9 recording lacked the necessary warmup data and was excluded.
  6. An October 14 partial-session comparison did not establish a reliable improvement.

Those are not footnotes. They define how much the experiment can teach us.

The research decision

The source report concluded that strategy-v2026.10.15 should remain unchanged. The one-minute filter should be monitored in shadow mode across a larger cohort of complete, replay-valid sessions.

Before promotion, the system would need to show stable, cost-aware net expectancy, acceptable drawdown, limited rejection of profitable opportunities, meaningful setup/regime coverage and no dependence on a threshold selected to fit this single date.

That is the practice we want the TradeStack Journal to reflect: publish the hypothesis, show the evidence, disclose the uncertainty, and keep testing.

For the detailed figures, methodology and verification scope, read the October 16 research case study. For the mechanisms behind replay, see our developer documentation.

Evidence of a smaller loss is evidence that one replay path differed. It is not yet evidence that a trading system has an edge.

Research, not a trade recommendation.

TradeStack documents paper-trading engineering and experiments. Performance results are only meaningful with their source, sample, fees, and data limitations.

CONTINUE EXPLORING

How the system fits together.

View the architecture ↗