TRADESTACK / RESEARCH CASE STUDY 001

A smaller loss
is not a proven edge.

We tested whether additional entry-quality checks could change paper-replay outcomes. The recorded comparison improved one partial session, but it did not produce a profitable result or justify changing production strategy rules.

16 OCT 202609:15–11:40 ISTCONTROLLED REPLAYLIMITED EVIDENCE
RESEARCH ONLY

Repository-documented comparison. Arithmetic independently checked; original market recording and replay were not rerun for publication.

01 / THE QUESTION

Would a stricter entry gate
reduce poor trades?

Instead of changing stop-loss rules after a loss, the experiment inspected whether completed candle information available at decision time could improve entry selection.

BASELINE

Existing strategy

Replay the documented frozen strategy without an additional quality gate.

VARIANT A

Immediate 1-minute quality

Reject a directional candidate when completed 1-minute data shows adverse EMA5/MACD momentum or a two-candle exhaustion pattern.

VARIANT B

Completed 5-minute alignment

Require the latest completed 5-minute candle direction and its close relative to EMA21 to agree with the candidate.

The experimental code used only information already completed at the decision timestamp. The report says it temporarily patched the analyzer inside the replay research process; it did not alter the configured paper-trading strategy.
02 / REPORTED RESULTS

One filter helped.
None made money.

All three variants were evaluated over the same recorded October 16 window. These values come from the internal experiment report, not audited live-account brokerage statements.

MARKET CALLBACK EVENTS57,579Reported recorded callbacks
ANALYZER EVALUATIONS380During the experiment window
CLOSED TRADES / VARIANT2Far too few for statistical inference
VariantClosed tradesWin rateGross P&LNet P&L after costsProfit factor
Existing baseline20%−₹1,240.50−₹1,459.770.00
Immediate 1-minute quality250%−₹273.00−₹491.050.51
Completed 5-minute alignment20%−₹1,240.50−₹1,459.770.00
Net loss magnitudeSmaller bar means smaller reported loss, not profit
Baseline
−₹1,459.77
1-min gate
−₹491.05
5-min gate
−₹1,459.77
CALCULATED FROM REPORTED RESULTS₹968.72

Improvement in net P&L from baseline to the immediate 1-minute variant. The outcome remained negative ₹491.05. The difference is a path-dependent replay comparison, not a forecast.

03 / WHY RESULTS CHANGED

Changing one decision
changes the trade path.

Different entries can change future cooldowns, open-position state, capital sizing and drawdown restrictions. Simply deleting losing trades from history would not recreate this counterfactual.

01

Original signals reviewed

For October 16, the report identifies rejected replay candidates at 09:33, 09:58 and 10:56 under the immediate 1-minute gate.

02

Later entries changed

The altered sequence subsequently produced different entries at 09:46 and 11:07 due to the evolving session state.

03

Five-minute confirmation did not help

The completed 5-minute gate still allowed the early losing positions; its reported after-cost outcome matched the baseline in this window.

04 / LIMITATIONS FIRST

What this experiment
does not prove.

The report's limitations are as important as its numerical result. These findings are exploratory, not a release recommendation.

Only partial sessions

Two partial sessions were usable in the cross-session investigation, with just two closed trades per October 16 variant. This cannot establish a robust expectancy estimate.

Replay does not fully match paper

The report says October 16 paper and replay outcomes were not identical. The controlled comparison supports hypothesis exploration, not a certified trade-by-trade reconstruction.

Some fields are not reliable yet

Replay trade setups were labeled UNKNOWN, preventing credible setup-level attribution from those rows.

Net results have fee uncertainty

Local fee estimates did not fully match broker estimates. The same cost model was used across variants, but absolute net P&L retains uncertainty.

An additional October 9 window had missing warmup data and no reconstructable trades; it was excluded. The October 14 partial window did not establish a reliable benefit.
05 / EVIDENCE STANDARD

What was checked.
What still needs proof.

We distinguish review of a source report and code from independently reproducing historical execution against the original private recording.

✓

Experiment report inspected

Reviewed the internal reports/research/immediate_1m_gate/FINDINGS.md study, including cohort scope, all three result rows, exclusions and limitations.

✓

Reported arithmetic cross-checked

Baseline implied costs: ₹219.27. Immediate-filter implied costs: ₹218.05. Net difference: +₹968.72. Both remain negative.

✓

Relevant engine implementation reviewed

Replayed data handling, session manifest validation, virtual clock, fill simulator tests and research workflows exist in the private Python codebase.

!

Historical market recording not independently rerun

The private October 16 raw recording was not processed afresh for this public case study. Numerical outputs remain attributed to the repository research report.

!

Broad independent replication not complete

Additional complete sessions, fresh fee reconciliation, paper/replay parity and an out-of-sample cohort are needed before strategy promotion.

06 / THE DECISION

Keep the hypothesis.
Do not promote the filter.

The internal report explicitly recommended leaving strategy-v2026.10.15 unchanged, collecting a larger paper-session cohort and monitoring the immediate 1-minute quality check in shadow mode. A better result across two trades is not evidence of repeatable profitability.