This is the empirical follow-up to “Which Strategies Can Survive XYZ Markets?”. The earlier essay identified possible improvements. This investigation actually compares them, rather than ending with a list of experiments someone should perform. We examined recorded, completed Forward opportunities under causally ordered strategy-specific market gates, paired original and Smart Entry/Exit positions at the same entry, and checked whether attractive strategy–regime interactions persisted across different dates.
XYZ is not intended to predict the next direction of the market. It classifies how the market has evolved and its currently observed fifteen-minute L1 state. That state is used to decide which trading strategies are appropriate to operate now. The evaluation in this essay therefore concerns the economic performance of strategy × observed market state, not forecast accuracy for future broad-market returns.
The principal result is uncomfortable but useful: a generic rule that permits strategies in historically profitable XYZ states failed on subsequent Forward dates. Yet Smart Entry selection separated better and worse opportunities within the very same XYZ states. The logical next improvement is not to discard XYZ, but to connect strategy-specific regime suitability, entry quality inside the regime and executable exit economics.
1. Four experiments actually completed
| Evaluation | What we did | What the comparison can tell us |
|---|---|---|
| Walk-forward strategy–regime gate | Determine state eligibility exclusively from trades already closed before a test date, then apply it to that day's Forward entries | Whether a historically good strategy-state combination retains its advantage later |
| Alternative present-state definitions | On identical trade candidates, compare Sliding15 XYZ, XY, Z, fixed fifteen-minute XYZ, rolling XYZ, marginal 1H and L1 plus 1H Z | Which way of describing the current state is most useful for strategy eligibility |
| Matched Smart Entry/Exit variants | Match the same instrument, an entry within fifteen seconds and entry price within 0.2%; verify common signal IDs where available | Whether improvement reflects selection of better entries or a superior exit policy |
| Date-by-date strategy–state review | Break promising and failing regime cells into their constituent active dates | Whether the apparent improvement is persistent or depends on one exceptional date |
At 14:26 KST on October 12, 2026, the Analysis Server contained 117,514 qualifying completed Forward trades. Of these, 117,383 had complete causal state tags for the common-method comparison; 131 lacked a required tag. There were 71 reconciled closed Actual trades, of which only 33 had valid linked entry-time L1 states.
The primary out-of-time evaluation covered the three completed test dates October 9–11. Each day's eligibility map was constructed using only trades that had already exited before that day began. October 12 was still a partial market day and is reported separately.
This is a read-only replay of filtering decisions on already recorded paper trades, not an actual rerun of exchange orders or a counterfactual portfolio in which freed capital enters different trades. We do not treat overlapping virtual strategies as independent account capital.
2. Experiment A: permitting historically positive XYZ cells did not produce positive prospective returns
We committed to three simple rules before evaluating their outcomes.
Exploratory allowlist: at least ten past completed positions for a given strategy × state, spread over two days, with a positive mean net return and PF at least 1.0.
Strict allowlist: at least twenty past positions over three dates, with positive mean return and PF at least 1.1.
Adverse-state veto: where at least fifteen positions over two days averaged −0.5% or worse and PF was below 0.85, suppress new entries of that strategy in that state.
The held-out, completed October 9–11 trades yielded:
| Rule applied to later entries | Retained N | Coverage | Retained mean net | PF |
|---|---|---|---|---|
| All recorded trade opportunities (baseline) | 63,900 | 100% | −0.415% | 0.325 |
| Sliding15 XYZ exploratory eligibility | 989 | 1.55% | −0.460% | 0.267 |
| Sliding15 XYZ strict eligibility | 74 | 0.12% | −0.326% | 0.402 |
| Sliding15 XY eligibility | 933 | 1.46% | −0.410% | 0.521 |
| Sliding15 Z eligibility | 852 | 1.33% | −0.336% | 0.553 |
| Fixed15 XYZ eligibility | 1,061 | 1.66% | −0.454% | 0.280 |
| Rolling XYZ eligibility | 1,215 | 1.90% | −0.469% | 0.241 |
| 1H XYZ eligibility | 632 | 0.99% | −0.430% | 0.338 |
| L1 XYZ + 1H Z eligibility | 149 | 0.23% | −0.462% | 0.238 |
| Sliding15 XYZ adverse-state veto | 61,653 | 96.48% | −0.413% | 0.327 |
The important result is not merely that some filtered average differs from baseline. No evaluated simple eligibility definition created positive net expectancy on the complete test dates. The default Sliding15 allowlist removed 98.45% of opportunities yet made the remaining average worse.
The strict threshold reduced average loss but admitted only 74 trades, or 0.12% of the original cohort, and still lost money. Not a single Core or ML opportunity met that strict per-cell eligibility history. In the broad common test cohort, the ML group averaged approximately +0.068% per trade across 223 positions; a sparse-cell gate that excludes every such opportunity has a measurable opportunity cost.
The adverse-state veto was not transformative either. It excluded 2,247 signals: 1,691 eventual losers but also 556 eventual winners. The retained mean improved only from roughly −0.415% to −0.413%, while PF remained far below 1.
Why a single good partial day is misleading
| Test date | Unrestricted mean net | Sliding15 exploratory mean |
|---|---|---|
| October 9 | −0.379% | −0.436% |
| October 10 | −0.375% | −0.450% |
| October 11 | −0.482% | −0.490% |
| October 12, partial day | −0.367% | −0.235% |
The gate underperformed on every completed test date yet looked relatively better during the unfinished October 12 window. This directly demonstrates why selecting a regime rule from whichever recent day happened to look favorable is not strategy research.
Decision: retain XYZ80 and the operational Sliding15 current-state classifier. Reject automatic whole-estate permission maps based solely on the historical per-cell average or hit rate. That is a negative result for the particular tested rules, not evidence that correctly applied XYZ strategy eligibility has no value.
3. Why the simple gates failed
Three mechanisms emerge from the experiment.
The eighty-state taxonomy divides a short history into sparse strategy-specific cells. The strict L1 gate admitted only 74 trades; strict L1 plus 1H Z admitted zero. More categories do not create more independent evidence, and a complicated L2–L3 combination can easily turn nearly every state into “unknown.”
Trading rows overstate independent opportunities. The 63,900 Forward observations on the three completed test dates corresponded to 11,325 distinct instrument-by-entry-minute events. Many Scalp variants react to the very same price shock. A hundred similar rows do not demonstrate a hundred independent confirmations of a regime.
A state classification cannot replace the quality of a strategy's entry. Within the same market state, one signal may capture an early sustainable expansion and another may buy after the move is nearly exhausted. This distinction is visible in the matched Smart Entry evidence below.
An economically useful strategy-state gate must therefore report avoided losing opportunities, missed winning opportunities, net returns on retained positions, execution costs, risk and any effect on finite capital allocation. Merely decreasing trade count is not success.
4. Experiment B: Runner Elevated Liquidity improved primarily by selecting better entries
The most consequential successful signal comes from Runner Elevated Liquidity. We linked 42 original and Smart Entry trades with the same instrument, near-identical entry time and price, and matching signal IDs.
| Runner Forward cohort | N | Win rate | Mean net |
|---|---|---|---|
| All baseline Runner trades | 113 | 36.3% | −0.257% |
| Original trades also represented by the same Smart Entry signals | 42 | 54.8% | +0.687% |
| Original trades lacking a corresponding Smart Entry record | 71 | 25.4% | −0.815% |
| The corresponding 42 Smart Entry trade outcomes | 42 | 54.8% | +0.684% |
| Entire Smart Entry variant | 97 | 44.3% | +0.174% |
The two original-strategy cohorts were separated by +1.503 percentage points of mean net return. The linked cohort performed better on five of six active days. A date-resampled sensitivity interval for the descriptive difference was roughly +0.456 to +2.699 points. This does not make selection randomized or guarantee performance in untouched markets.
Crucially, the same 42 entry signals did not show an economically important exit difference: original and Smart Entry outcomes differed by approximately −0.004 percentage points on average.
Interpretation: the evidence points toward entry opportunity selection, not a magical profit-taking algorithm.
The advantage existed even within the same XYZ market state
| Observed L1 regime, same Runner parent | Linked Smart Entry signal in original | Unlinked original signal |
|---|---|---|
| DEAD_CAT_BOUNCE · FLAT | 5 trades, +1.128% | 5 trades, −2.385% |
| SELECTIVE_INFLOW · STRONG_UP | 5 trades, +1.301% | 3 trades, −2.094% |
These cells are too small to define a Live rule on their own. Nevertheless, they explain the product's intended architecture better than a global “good regime” list. Excluding all Runner trades in DEAD_CAT_BOUNCE · FLAT would also exclude the profitable linked signals. Allowing them all would retain weak ones.
Improvement decision: preserve the existing Runner Smart Entry logic as a research benchmark. For every original parent signal, log observed XYZ, instrument-level liquidity, executable spread and depth, ATR-normalized entry extension, whether Smart Entry selected it, and the actual reason for selection or absence. An unlinked original position is not automatically proven to have been explicitly rejected; it may have been unavailable because of capacity, timing or concurrent positions.
The next controlled comparison is XYZ alone versus Smart Entry alone versus XYZ plus Smart Entry, on the same parent opportunities with all rejected candidates shadow-tracked.
5. Experiment C: Fibonacci Smart Entry also separated better and worse opportunities
| Fibonacci Forward cohort | N | Win rate | Mean net |
|---|---|---|---|
| Entire Fibonacci baseline | 49 | 32.7% | −0.569% |
| Baseline signals also matched to Smart Entry | 11 | 54.5% | +0.787% |
| Baseline signals without matching Smart Entry record | 38 | 26.3% | −0.961% |
| Smart Entry results on the same eleven signals | 11 | 54.5% | +0.894% |
| Entire Smart Entry variant | 46 | 43.5% | +0.119% |
The original-strategy cohorts differed by +1.748 percentage points. Linked entries were relatively stronger on five of six days. The date-resampled sensitivity interval was broad, roughly +0.061 to +3.378 points, and only eleven original signals could be exactly matched.
On those same eleven entries, the improvement attributable to the recorded variant's different exit execution was only +0.107 points on average; its date-clustered interval crossed zero.
Improvement decision: the highest-priority Fibonacci question is no longer “Which new exit variant should be added?” It is what information at the entry time distinguished a sustainable retracement recovery from a failed rebound. Candidate explanatory signals include acceptance of the Fibonacci level, higher-horizon selling pressure, volume recovery, price extension and the exact bid/ask achievable after confirmation.
Keep the Smart Entry variant under prospective research. Eleven matched cases are not enough to increase funded exposure.
6. Experiment D: Ignition Persistent became less bad, not profitable
For Ignition Breakout 1H Persistent, 58 positions linked to Smart Entry B by identical signal IDs. Those matched original trades averaged −0.110%; 63 original trades without a corresponding Smart Entry record averaged −0.852%.
That is a +0.742-point descriptive separation, with a date-resampled interval of approximately +0.237 to +1.035 points. Applying the Smart Entry B variant to the 58 matched opportunities produced −0.054% mean net, or about +0.056 points over the matched original outcome.
The essential result is that the improved trade set still loses money. A strategy that becomes less unprofitable is not yet an investable strategy.
A concrete regime-specific failure candidate
When Persistent Ignition operated in SELECTIVE_INFLOW · STRONG_UP, the observation was eleven trades across five active dates, zero net winners and −1.093% mean net return.
That does not imply XYZ forecast a market decline. The label simply described a present market condition in which this particular strategy's recorded entries all lost.
Improvement decision: freeze this specific strategy × XYZ cell as a candidate for shadow veto testing, while continuing to record the original strategy's hypothetical outcomes. Measure losses it would avoid and profitable opportunities it would miss on later dates. The eleven observations that motivated the rule must not be reused as fresh validation. Do not archive the entire strategy or prohibit every other strategy from operating in this state.
7. Experiment E: the apparent Predictive B/C superiority shrank in entry-matched comparisons
At the whole-variant level, Adaptive Predictive Model's baseline averaged about +0.323%, Exit B +0.625%, and Exit C +0.391%. It is tempting to declare B the universal winner, but these variants did not necessarily trade identical opportunity sets.
We found 64 same-instrument, same-entry-time, same-price B/C pairs.
| Exit on a matched opportunity | Matched trades | Mean net |
|---|---|---|
| B | 64 | +0.625% |
| C | 64 | +0.543% |
| B minus C | 64 | +0.082 pp |
Forty-seven pairs had identical returns. B beat C in ten cases; C beat B in seven. The date-resampled interval for B's advantage was approximately −0.017 to +0.172 percentage points. It included zero.
For the 33 matches between the ordinary Predictive baseline and B, baseline averaged +0.093%, compared with +0.204% for B. The matched difference of +0.110 points is much smaller than the +0.302-point difference between the unpaired whole-variant averages.
These B/C matches establish economic similarity of entry time and price, not verified identity of every parent signal ID. Fixing that lineage is a prerequisite to a definitive exit-only causal study.
Improvement decision: retain both B and C as distinct Forward experiments. Compare fully identical parent signals to identify when earlier profit locking beats a longer hold, and when the longer hold captures upside B sacrifices. Do not delete C or generalize B to unrelated strategy families from these results.
8. Some strategy–XYZ relationships look worth preserving; others are specifically dangerous
We decomposed noteworthy observations by active date rather than looking at the aggregate average alone.
| Strategy × currently observed L1 state | Observed return | What the individual dates show | Research classification |
|---|---|---|---|
| Predictive c0005 × SELECTIVE_INFLOW · UP | 12 trades, 75% WR, +1.994% | Five dates; October 11 averaged −0.1% | Candidate eligible cell, not a funded gate |
| Runner × SELECTIVE_INFLOW · UP | 11 trades, +1.434% | Five dates; October 12 averaged −0.194% | Test together with the Smart Entry selection |
| Relative Strength × SELECTIVE_INFLOW · STRONG_UP | 10 trades, +1.341% | Negative days also occurred | Candidate, not an unconditional permission |
| Persistent Ignition × SELECTIVE_INFLOW · STRONG_UP | 11 trades, −1.093%, 0% WR | Lost on all five active dates | High-priority strategy-specific veto candidate |
| Liquidity/Inflow × SELECTIVE_INFLOW · FLAT | 27 trades, −1.661% | Repeated large losses over eight dates | Candidate adverse context for this strategy |
| Liquidity/Inflow × SELECTIVE_INFLOW · STRONG_UP | 15 trades, +0.367% | One day +7.974%, another day −3.908% | Unstable; no eligibility promotion |
This comparison prevents two common errors. A positive five-day conditional average may still be fragile; and an adverse state for one strategy does not imply a universally adverse state for every strategy.
For Liquidity/Inflow, the difference between FLAT and STRONG_UP within SELECTIVE_INFLOW is worth investigating, but the highly volatile daily results suggest actual instrument-level buying pressure and entry maturity, not just the Z sign, may be decisive.
For Predictive c0005, the twelve-trade +1.994% observation is a reason to freeze that state hypothesis and evaluate entirely later Forward dates. It is not a sufficient basis for immediate funded activation.
9. Scalp and higher horizons received their own negative results
On the three complete test dates, Scalp recorded 59,547 opportunities averaging −0.414%. The basic Sliding15 XYZ eligibility rule retained 954 (1.60%), but those selected trades averaged −0.444%. A generic per-cell Scalp filter did not demonstrate an improvement.
A few individual selected Scalp subsets were marginally positive, but typical counts were only about ten trades across two days. Because multiple variants trade the same instrument-minute, these are far from independent confirmations of an edge.
Decision for Scalp: first collapse variants into underlying parent signal events, preserve each child variant's incremental information, and assess filters on the same executable market opportunities with finite capital and orderbook costs. Do not use a sparse per-cell hit rate as an automatic Live promotion signal.
For higher horizons, 1H XYZ admitted 632 trades averaging −0.430%; L1 XYZ plus 1H Z admitted 149 averaging −0.462%. The strict combined condition admitted no trades.
Decision for L2: keep operational L1 at fifteen minutes. Introduce 1H/4H/1D/7D as selective current-market-context refinements only for a strategy where later independent evidence shows better separation of eligible and ineligible conditions. A more elaborate state name is not itself an improvement.
10. Actual execution is the remaining constraint on claims of real profitability
The exchange-reconciled Actual sample now contains 71 closures, but only 33 with usable L1 classifications at entry. That is not enough to declare Live profitability for the new strategy–regime combinations, especially among Scalp, Adaptive and V8/V10/V11 research descendants.
Earlier matched Upbit examples, such as CPOOL and APT, showed identical simulated and executed entry prices but less favorable exchange exit fills. A correct current market-state label cannot compensate for a fictitious exit price or missing trading costs.
A reliable decision record must connect parent Signal ID → time-available features → observed XYZ state → entry selection/denial → child exit experiment → order and fill → realized KRW net PnL.
Two additional evidence-quality populations remain important: 57,334 previously identified incomplete short-timeframe Feature Mart records, and 131 completed Forward trades lacking some common classification tags in this comparison. They must remain visible as quality categories rather than being silently counted as valid losses or deleted from denominators without explanation.
Decision: the positive Forward observations in this article are grounds for targeted research, not automatic new Live permissions. Incomplete Actual context must be recovered from genuine decision-time evidence, never manufactured by using the later current regime.
11. Improvement decisions and concrete implementation acceptance criteria
The following are decisions from the completed read-only research and specifications for the next implementation. They do not claim that live eligibility settings, strategy weights, orders or capital allocation have already been modified.
| Priority | Target | Verified finding | Improvement decision | Required evidence before adoption |
|---|---|---|---|---|
| P0 | Signal provenance across every strategy | Runner/Fibonacci Smart Entry could match signal IDs; Predictive B/C could only match time and price | Require common parent Signal ID, child ID and explicit accept/reject reason | Every future parent opportunity traceable through all variant evaluations |
| P0 | Actual and Feature Mart integrity | 71 Actual closures, 33 with L1; 131 Forward missing common tags | Restore reliable event-time feature, state and exchange-fill joins | No guessed regimes or non-causal features in trading evidence |
| P1 | Runner Smart Entry | 42 linked signals +0.687%; 71 unlinked originals −0.815% | Preserve the selection layer, benchmark XYZ alone / entry alone / both | At least ten new trading days and 100 same-signal pairs with positive net execution-aware value, including missed winners |
| P1 | Fibonacci Smart Entry | Eleven linked signals +0.787% versus 38 unlinked −0.961% | Retain the candidate; identify successful recovery conditions | At least ten new trading days and 50 paired parent signals, then confirm net value |
| P1 | Persistent Ignition | Entry filtering reduces losses but leaves negative expectancy | Maintain research; no expansion of funded risk | Positive later-period net expectancy, not merely better relative rank |
| P1 | Persistent × SELECTIVE_INFLOW · STRONG_UP | Eleven of eleven trades lost over five dates | Freeze a strategy-only shadow veto hypothesis | At least ten later distinct dates evaluating losses avoided and winners missed |
| P1 | Predictive Exit B/C | B +0.082 pp on 64 price/time pairs, not decisively superior | Keep parallel variants and repair common signal lineage | At least fifteen new days, 100+ identical-parent entries, compare MFE, downside, fees |
| P1 | Predictive c0005 state candidate | Twelve SELECTIVE_INFLOW · UP trades averaged +1.994% | Freeze the conditional eligibility hypothesis | Reproduce on uninspected new Forward days and exchange-compatible fills |
| P2 | Scalp | Only 954 of 59,547 tested entries retained; still −0.444% | Rebuild analysis around independent parent market events | Positive realized net results after correlated exposure and execution costs |
| P2 | Optional higher-horizon gates | 1H and L1×1H alternatives sparse and negative | Do not deploy a general L2 gate | Demonstrate incremental conditional eligibility value over L1 for that strategy |
The operating rule that follows from the evidence
Every strategy should have a separate assessment in every observable XYZ80 state: eligible, ineligible, or insufficient evidence. A failing state should restrict that particular strategy, not automatically archive the entire strategy or prohibit every other strategy in the market.
Unknown means unknown, not zero wins and not permission to trade with real money. Historical research Holdout cannot become new Forward evidence merely by being displayed on a different page.
Rejected parent signals must continue to receive shadow outcomes. Otherwise, a gate that suppresses trade count can appear attractive while excluding the most profitable opportunities.
Change one mechanism at a time. For Runner, evaluate entry selection; for Fibonacci, recovery confirmation; for Persistent, a specific strategy-state exclusion; for Predictive, matched exit timing. Combining all changes at once would make the next performance shift impossible to explain.
12. Final decision: what survived empirical comparison
The improvement that did not survive: mechanically permitting all strategies only in their historically positive XYZ categories. It did not deliver profitable Forward results on the three completed test dates, and switching from Sliding15 to Fixed, Rolling or a sparse higher-horizon combination did not rescue it.
The improvements worth developing: Runner and Fibonacci Smart Entry showed compelling selection differences between matching and unmatching original opportunities. Predictive Exit B and C remained promising but the true matched difference was smaller than their overall averages suggested. Persistent Ignition supplied a specific strategy-state failure hypothesis rather than a reason to eliminate the whole strategy family.
The architecture is therefore straightforward:
XYZ describes how the market arrived at its present condition. Actual strategy-by-state evidence determines which strategies may act in that condition. Smart Entry then decides which opportunities within the allowed condition deserve a real order, and execution determines the profit actually realized.
This is how a descriptive XYZ taxonomy becomes a strategy eligibility system without being mistaken for a directional market forecast. The next achievement to seek is not a prettier historical win-rate map but a reproducible improvement in net executable outcomes on genuinely new opportunities.