SMC Inverse Fair Value Gap: An Empirical Test

Test period: January 2012 – October 2026 · Instruments: EURUSD, GBPUSD, USDJPY, AUDUSD, XAUUSD, BTCUSD (from May 2017), US100 (M1, bid) · Out-of-sample block: 2012–2018 · Timeframes: M15, H1, H4 · Published: November 2026

near edge (gap low): short limit entry MT (50% of the gap) far edge (gap high): stop bullish FVG: low of candle 3 > high of candle 1 close below the gap = inversion first return
Fig. 1 — The inverse fair value gap (iFVG) definition used in this study (bearish case; bullish is mirrored). A bullish fair value gap is left by three candles; a later candle closes below the gap instead of respecting it. The original gap now acts as resistance: a sell limit is placed at its near edge (the gap low) for the first return, with the stop at the far edge (the gap high).

1. Objective

A fair value gap (FVG) is a three-candle pattern in which the first and third candles do not overlap, leaving a price range the market moved through quickly. The standard idea is that price returns to fill part of the gap and then continues in the original direction. The inverse fair value gap is a gap that fails: price closes through it instead of respecting it. The gap is then said to “invert”, so that a failed bullish gap becomes resistance and a failed bearish gap becomes support. Traders sell the first return to a failed bullish gap (and buy the first return to a failed bearish one), with the stop beyond the far edge of the gap.

The concept is also usually described with a few qualifiers: it works better when the inversion is forceful, when the gap is fresh, and when it forms in the premium (for sells) or the discount (for buys) of the recent range. This study turns these statements into fixed, mechanical rules and tests them against a random baseline across seven instruments, three timeframes and nearly 15 years of one-minute data, in the same format as the Order Blocks study.

1.1 Claims tested

#ClaimSection
①A limit order on the first return to an iFVG beats random entries and has a positive expectancy after costs3.1
②The inversion carries information: the failed gap is a better level than the same gap traded the ordinary way (first return, original direction)3.2
③Refinements help: entering at the 50% level of the gap, or placing the stop 0.25 ATR beyond the far edge3.3
④Quality matters: a more forceful inversion candle, a larger gap, or a fresher gap (inverted soon after it formed) works better3.4
⑤Context matters: an iFVG works better in premium (sells) or discount (buys) of the recent range3.5

2. Methodology

2.1 Data

  • Source: Dukascopy, M1 bid (via the dukascopy-node CLI). BTCUSD is available from May 2017 only
  • Range: 2012-01-01 to 2026-10-31, about 4.2–5.5 million one-minute bars per instrument after removing flat (OHLC-equal) bars
  • Candles: M15 and H1 on clock boundaries; the 4-hour candles are NY-aligned (opening at 18:00, 22:00, 02:00, 06:00, 10:00 and 14:00 New York time, with full daylight-saving rules), as in the Order Blocks study
  • Out-of-sample split: all definitions and parameters were written down before the first run. Results were examined on 2019–2026 first; the 2012–2018 block was then run once with no changes, so it acts as an out-of-sample check. Every result is reported for both blocks.

2.2 iFVG definition (fixed before testing)

  • FVG: three consecutive candles. Bullish if the low of the third candle is above the high of the first; the gap is the range between them (bearish mirrored). Gap size of at least 0.1 × ATR(14). The gap exists from the close of the third candle
  • Inversion: within 24 candles after the gap is complete, a candle closes beyond the far side of the gap (below its low for a bullish gap, above its high for a bearish one). That is the first close through the gap; wicks into the gap do not count. A failed bullish gap gives a sell setup (bearish iFVG); a failed bearish gap gives a buy setup
  • Zone: the original gap. Near edge = the edge first met on the return (gap low for a sell, gap high for a buy); mean threshold (MT) = the middle of the gap; far edge = the opposite edge
  • No look-ahead: the setup is live only from the close of the inverting candle. Entry orders are placed after that candle closes
  • First return: the first touch of the entry level within 48 candles, found on M1 data. If the touching M1 bar opens beyond the limit price (a gap into the zone), the fill is that bar’s open
  • Control (plain FVG): every FVG, traded the ordinary way: a limit order at the near edge on the first return in the original direction, stop at the far edge, same 48-candle window and same exit rules

The rules found 336,126 inversions among 524,853 fair value gaps across all instruments and timeframes (198,731 of 309,514 in 2019–2026 and 137,395 of 215,339 in 2012–2018), so about 64% of gaps fail inside 24 candles. Price returned to the near edge of the inverted zone within 48 candles in 91–93% of cases.

2.3 Random baseline

Every trade is paired with ten random entries in the same direction, at the same minute of the day (UTC), on other days within ±60 days, with the same stop distance. Random entries are market orders at that minute’s open and follow exactly the same exit rules. The quantity reported throughout is the edge: the real win rate minus the random win rate, in percentage points (pp). Standard errors and z-values use a month-block bootstrap, because events overlap in time.

2.4 Outcomes and costs

  • Fixed targets of 1R and 2R; whichever of target or stop is hit first on M1 data, within three days. Trades still open after three days are closed at market
  • If the entry M1 bar itself already touches the stop, the trade is counted as a loss; if target and stop are both hit inside the same later M1 bar, the trade is excluded as unresolved
  • Minimum stop distance: 10 × the round-trip cost, so costs never exceed 0.1R. This removes most small zones: the share of iFVG setups that remain is 4% on M15, 13% on H1 and 35% on H4 in 2019–2026, and 0.8%, 7% and 25% in 2012–2018, when volatility was lower relative to costs. Results therefore describe the larger gaps
  • Net expectancy is reported after an approximate prop-firm cost (spread plus commission, round trip, October 2026):
EURUSDGBPUSDUSDJPYAUDUSDXAUUSDBTCUSDUS100
1.4 pips2.2 pips2.5 pips1.6 pips$0.58$301.5 pts

3. Results

3.1 Claim ① — Does the first return to an iFVG beat random?

Win rate minus random (pp) -8 -6 -4 -2 0 +2 +4 +6 +8 Win rate minus random (pp) M15 · 2012–2018: -0.49 pp (95% CI ±3.91) M15 · 2019–2026: +0.66 pp (95% CI ±1.45) M15 H1 · 2012–2018: +2.47 pp (95% CI ±2.73) H1 · 2019–2026: +1.95 pp (95% CI ±1.79) H1 H4 · 2012–2018: +0.21 pp (95% CI ±2.21) H4 · 2019–2026: +0.01 pp (95% CI ±2.06) H4 all · 2012–2018: +0.89 pp (95% CI ±1.81) all · 2019–2026: +0.93 pp (95% CI ±1.06) All timeframes 2012–2018 (out-of-sample) 2019–2026
Fig. 2 — Win rate at 1R minus the random baseline, iFVG limit entry at the near edge (whiskers = 95% CI). The pooled edge is small and positive in both blocks, but it is not significant in either, and the timeframes disagree with each other.
TFPeriodTargetnWin rateRandomEdge, pp (z)Gross exp.Net exp.
M152012–20181R89848.2%48.7%-0.49 (z -0.2)-0.025R-0.101R
M152012–20182R89831.5%30.9%+0.66 (z +0.4)-0.017R-0.094R
M152019–20261R5,96050.1%49.4%+0.66 (z +0.9)+0.002R-0.064R
M152019–20262R5,96833.7%32.9%+0.84 (z +1.3)+0.018R-0.048R
H12012–20181R1,55949.8%47.3%+2.47 (z +1.8)+0.016R-0.055R
H12012–20182R1,56031.0%28.8%+2.18 (z +1.6)+0.005R-0.066R
H12019–20261R4,57051.1%49.2%+1.95 (z +2.1)+0.029R-0.031R
H12019–20262R4,57133.3%32.2%+1.12 (z +1.4)+0.036R-0.023R
H42012–20181R1,80447.2%47.0%+0.21 (z +0.2)-0.010R-0.073R
H42012–20182R1,80429.4%27.7%+1.70 (z +1.7)+0.017R-0.047R
H42019–20261R3,32348.0%48.0%+0.01 (z +0.0)-0.011R-0.065R
H42019–20262R3,32329.6%29.5%+0.12 (z +0.1)+0.003R-0.051R
All TFs2012–20181R4,26148.3%47.5%+0.89 (z +1.0)-0.003R-0.072R
All TFs2012–20182R4,26230.4%28.8%+1.66 (z +2.0)+0.005R-0.064R
All TFs2019–20261R13,85349.9%49.0%+0.93 (z +1.7)+0.008R-0.053R
All TFs2019–20262R13,86232.6%31.9%+0.76 (z +1.4)+0.020R-0.041R
All TFsBoth blocks1R18,11449.5%48.6%+0.92 (z +2.2)+0.005R-0.058R
All TFsBoth blocks2R18,12432.1%31.1%+0.97 (z +2.2)+0.017R-0.046R

Limit entry at the near edge, stop at the far edge. Net expectancy is per trade, in R, after the cost table in Section 2.4.

Pooled over all timeframes, the iFVG wins 49.9% of 1R trades against 49.0% for random entries with the same stop in 2019–2026 (edge +0.93 pp, z +1.7), and 48.3% against 47.5% in 2012–2018 (edge +0.89 pp, z +1.0). Both blocks point to the same small positive edge of about one percentage point, but neither reaches z = 2 on its own (pooled over both blocks: +0.92 pp, z +2.2), and at 2R the result is +0.76 pp (z +1.4) and +1.66 pp (z +2.0). Before costs the setup earns +0.008R per trade in 2019–2026 and -0.003R in 2012–2018; after costs, -0.053R and -0.072R. H1 is the most consistent timeframe (+2.47 pp and +1.95 pp at 1R, z +1.8 and +2.1), but it is one of twelve comparisons in the table and is still negative after costs (-0.055R and -0.031R). A win-rate edge of one point is worth about 0.02R at 1R, while the assumed cost is 0.06–0.07R per trade, so no timeframe or block comes out positive after costs.

3.2 Claim ② — Does the inversion carry information?

A failed gap is the same price range as an ordinary gap; only the direction of the trade is reversed after the close through it. If the failure were informative, the iFVG should beat the ordinary first return to the same kind of gap.

iFVG minus plain FVG (pp) -8 -6 -4 -2 0 +2 +4 +6 +8 iFVG minus plain FVG (pp) M15 · 2012–2018: -1.05 pp (95% CI ±4.40) M15 · 2019–2026: -0.49 pp (95% CI ±1.74) M15 H1 · 2012–2018: +0.27 pp (95% CI ±3.16) H1 · 2019–2026: +0.41 pp (95% CI ±2.21) H1 H4 · 2012–2018: -0.93 pp (95% CI ±2.83) H4 · 2019–2026: -1.19 pp (95% CI ±2.50) H4 all · 2012–2018: -0.45 pp (95% CI ±2.15) all · 2019–2026: -0.36 pp (95% CI ±1.35) All timeframes 2012–2018 (out-of-sample) 2019–2026
Fig. 3 — Edge of the iFVG minus edge of the plain FVG first return, 1R (whiskers = 95% CI). No value is significant, and the sign changes between timeframes.
TFPeriodTargetn (iFVG / plain FVG)iFVG edge (pp)Plain FVG edge (pp)Difference (z)Net exp. (iFVG / plain FVG)
M152012–20181R898 / 3,316-0.49+0.56-1.05 (z -0.5)-0.101R / -0.059R
M152012–20182R898 / 3,317+0.66+0.61+0.05 (z +0.0)-0.094R / -0.062R
M152019–20261R5,960 / 12,583+0.66+1.16-0.49 (z -0.6)-0.064R / -0.055R
M152019–20262R5,968 / 12,589+0.84+0.68+0.15 (z +0.2)-0.048R / -0.062R
H12012–20181R1,559 / 3,800+2.47+2.20+0.27 (z +0.2)-0.055R / -0.029R
H12012–20182R1,560 / 3,802+2.18-0.35+2.53 (z +1.6)-0.066R / -0.069R
H12019–20261R4,570 / 8,903+1.95+1.53+0.41 (z +0.4)-0.031R / -0.038R
H12019–20262R4,571 / 8,911+1.12+1.45-0.33 (z -0.3)-0.023R / -0.021R
H42012–20181R1,804 / 3,358+0.21+1.14-0.93 (z -0.6)-0.073R / -0.036R
H42012–20182R1,804 / 3,360+1.70-0.24+1.95 (z +1.5)-0.047R / -0.053R
H42019–20261R3,323 / 5,861+0.01+1.20-1.19 (z -0.9)-0.065R / -0.042R
H42019–20262R3,323 / 5,863+0.12+0.52-0.40 (z -0.3)-0.051R / -0.045R
All2012–20181R4,261 / 10,474+0.89+1.34-0.45 (z -0.4)-0.072R / -0.041R
All2012–20182R4,262 / 10,479+1.66-0.01+1.67 (z +1.6)-0.064R / -0.062R
All2019–20261R13,853 / 27,347+0.93+1.29-0.36 (z -0.5)-0.053R / -0.047R
All2019–20262R13,862 / 27,363+0.76+0.90-0.14 (z -0.2)-0.041R / -0.045R

The plain FVG first return itself shows an edge of +1.29 pp in 2019–2026 and +1.34 pp in 2012–2018 at 1R, which matches the roughly +1.3 pp found for 4-hour FVGs in the Order Blocks study. The iFVG is at +0.93 pp and +0.89 pp. The difference at 1R is -0.36 pp (z -0.5) and -0.45 pp (z -0.4); at 2R it is -0.14 pp (z -0.2) and +1.67 pp (z +1.6). Trading the failure of a gap is statistically indistinguishable from trading the gap itself, and both lose money after costs (net -0.053R / -0.047R in 2019–2026).

3.3 Claim ③ — Entry and stop refinements

Entry / stopPeriodn1R: edge, pp (z)1R: net exp.2R: edge, pp (z)2R: net exp.
Near edge, stop at the far edge (baseline)2012–20184,261+0.89 (z +1.0)-0.072R+1.66 (z +2.0)-0.064R
Near edge, stop at the far edge (baseline)2019–202613,853+0.93 (z +1.7)-0.053R+0.76 (z +1.4)-0.041R
Near edge, stop 0.25 ATR beyond the far edge2012–20188,146+0.53 (z +0.8)-0.078R+1.18 (z +2.0)-0.064R
Near edge, stop 0.25 ATR beyond the far edge2019–202625,186+0.79 (z +2.1)-0.057R+0.83 (z +2.3)-0.045R
Mean threshold (50% of the gap), stop at the far edge2012–2018722-0.47 (z -0.2)-0.109R+0.57 (z +0.3)-0.095R
Mean threshold (50% of the gap), stop at the far edge2019–20264,214+1.63 (z +2.0)-0.041R+1.41 (z +1.9)-0.022R

All timeframes pooled.

  • Stop 0.25 ATR beyond the far edge. Wider stops let almost twice as many setups pass the minimum-stop rule (more events in the table), but the edge stays at +0.79 pp (2019–2026) and +0.53 pp (2012–2018), and net expectancy at 1R is -0.057R and -0.078R.
  • Entry at the mean threshold. Waiting for a deeper fill at the middle of the gap is filled much less often. It looks better in 2019–2026 (+1.63 pp, z +2.0) but is -0.47 pp (z -0.2) in 2012–2018 on 722 trades, and is negative after costs in both blocks.

3.4 Claim ④ — Forceful inversion, larger gap, fresher gap

AttributeBinn (2012–18 / 2019–26)Edge 2012–2018Edge 2019–2026Net exp. (2012–18 / 2019–26)
Inversion candle body (× ATR)<0.5930 / 3,124-0.05 (z -0.0)-0.42 (z -0.4)-0.089R / -0.069R
Inversion candle body (× ATR)0.5-1.01,443 / 4,884+0.13 (z +0.1)+0.97 (z +1.1)-0.094R / -0.051R
Inversion candle body (× ATR)1.0-1.5984 / 3,076+1.45 (z +1.0)+2.30 (z +2.1)-0.055R / -0.034R
Inversion candle body (× ATR)>=1.5904 / 2,769+2.47 (z +1.3)+0.86 (z +0.7)-0.039R / -0.060R
Gap size (× ATR)<0.2516 / 795—+0.37 (z +0.2)— / -0.099R
Gap size (× ATR)0.25-0.5331 / 3,102-1.99 (z -0.7)+1.03 (z +1.0)-0.139R / -0.059R
Gap size (× ATR)0.5-1.01,658 / 5,789+0.48 (z +0.3)+0.45 (z +0.7)-0.076R / -0.054R
Gap size (× ATR)>=1.02,256 / 4,167+1.60 (z +1.2)+1.63 (z +1.7)-0.060R / -0.038R
Candles between FVG and inversion1184 / 1,033-4.18 (z -1.0)-0.01 (z -0.0)-0.195R / -0.080R
Candles between FVG and inversion2-3542 / 2,489+0.31 (z +0.1)+1.22 (z +1.0)-0.091R / -0.049R
Candles between FVG and inversion4-81,339 / 4,296+1.55 (z +1.0)+1.43 (z +1.9)-0.057R / -0.044R
Candles between FVG and inversion9-242,196 / 6,035+1.06 (z +0.9)+0.62 (z +0.8)-0.067R / -0.056R
Middle-candle body of the FVG (× ATR)<0.5324 / 1,021+5.28 (z +1.4)+0.94 (z +0.5)-0.013R / -0.065R
Middle-candle body of the FVG (× ATR)0.5-1.0610 / 3,597+0.25 (z +0.1)+1.42 (z +1.5)-0.058R / -0.047R
Middle-candle body of the FVG (× ATR)1.0-1.51,082 / 4,079+0.78 (z +0.5)+0.37 (z +0.4)-0.076R / -0.069R
Middle-candle body of the FVG (× ATR)>=1.52,245 / 5,156+0.48 (z +0.4)+1.03 (z +1.4)-0.083R / -0.042R

iFVG, limit entry at the near edge, 1R, all timeframes pooled. Bins were set before the run.

None of the attributes sorts winners from losers in a way that holds in both blocks. The bins do not line up in order, and the best-looking cells in one block are flat or negative in the other. Fresher gaps are not better: the very freshest (inverted on the next candle) are the weakest bin in 2019–2026 and negative in 2012–2018, and the gaps that took 4–8 candles to invert are the best in both blocks, at about +1.4 to +1.5 pp, which is within noise.

3.5 Claim ⑤ — Premium and discount

For each setup, the middle of the gap is located within the high–low range of the previous 96 candles, from 0% (bottom) to 100% (top). For a sell the position is read as is, for a buy it is flipped, so that 100% always means “premium for a sell, discount for a buy”. The claim, as written before testing, was that setups in the favourable half (50–100%) beat those in the unfavourable half. The difference was +2.17 pp (z +1.2) in 2012–2018 and +1.07 pp (z +1.1) in 2019–2026, in the expected direction, but not significant.

Position of the gap in the rangen (2012–18 / 2019–26)Edge 2012–2018Edge 2019–2026Net exp. (2012–18 / 2019–26)
0-25%434 / 2,009-2.53 (z -1.0)+0.38 (z +0.3)-0.086R / -0.055R
25-50%1,055 / 3,362+0.31 (z +0.2)+0.21 (z +0.3)-0.077R / -0.048R
50-75%1,522 / 4,391-0.38 (z -0.3)+0.03 (z +0.0)-0.096R / -0.082R
75-100% (outer quartile)1,250 / 4,091+4.11 (z +2.2)+2.76 (z +4.0)-0.035R / -0.026R

iFVG, limit entry at the near edge, 1R, all timeframes pooled.

The quartile table does show one cell that stands out: setups in the outer quartile (75–100%) beat random by +2.76 pp (z +4.0) in 2019–2026 and by +4.11 pp (z +2.2) in 2012–2018, while the other three quartiles are within noise of zero. This is the only subgroup in the study that stands out clearly in both blocks. It was, however, found by looking at the bins, among about twenty (five attributes with four bins each), and the follow-up below was run after seeing it. It should be read as exploratory.

Win rate minus random (pp) -4 -2 0 +2 +4 +6 +8 +10 Win rate minus random (pp) iFVG · outer quartile · 2012–2018: +4.11 pp (95% CI ±3.75) iFVG · outer quartile · 2019–2026: +2.76 pp (95% CI ±1.36) iFVG outer quartile iFVG · the rest · 2012–2018: -0.45 pp (95% CI ±1.77) iFVG · the rest · 2019–2026: +0.16 pp (95% CI ±1.25) iFVG the rest plain FVG · outer quartile · 2012–2018: +5.14 pp (95% CI ±2.93) plain FVG · outer quartile · 2019–2026: +2.91 pp (95% CI ±1.73) plain FVG outer quartile plain FVG · the rest · 2012–2018: +0.94 pp (95% CI ±1.26) plain FVG · the rest · 2019–2026: +1.03 pp (95% CI ±0.88) plain FVG the rest 2012–2018 (out-of-sample) 2019–2026
Fig. 4 — Edge at 1R for setups in the outer quartile of the 96-candle range versus the rest, for iFVG and for the plain FVG first return (whiskers = 95% CI). The outer-quartile effect is as large for the plain FVG as for the iFVG.
SetPositionn (2012–18 / 2019–26)Edge 2012–2018Edge 2019–2026Net exp. (2012–18 / 2019–26)
iFVGOuter quartile (75–100%)1,250 / 4,091+4.11 (z +2.2)+2.76 (z +4.0)-0.035R / -0.026R
iFVGRest (0–75%)3,011 / 9,762-0.45 (z -0.5)+0.16 (z +0.3)-0.088R / -0.064R
plain FVGOuter quartile (75–100%)997 / 3,838+5.14 (z +3.4)+2.91 (z +3.3)+0.005R / -0.026R
plain FVGRest (0–75%)9,477 / 23,509+0.94 (z +1.5)+1.03 (z +2.3)-0.046R / -0.050R

The follow-up shows that the effect is a property of where the gap is, not of the inversion. The plain FVG first return, with no inversion, gains +2.91 pp (z +3.3) in 2019–2026 and +5.14 pp (z +3.4) in 2012–2018 in the same outer quartile, at least as much as the iFVG (+2.76 pp and +4.11 pp). This is the same pattern as in the Order Blocks study, where being at a local extreme accounted for the whole difference between an OB and an ordinary candle. The random baseline does not control for position in the range, so a trade placed at the edge of the range is compared with random entries from anywhere. After costs, the outer-quartile iFVG is still slightly negative (-0.026R in 2019–2026, -0.035R in 2012–2018), while the plain FVG in the same cell is -0.026R and +0.005R.

3.6 By instrument

Instrumentn (2012–18 / 2019–26)Edge 2012–2018Edge 2019–2026Net exp. 2012–2018Net exp. 2019–2026
EURUSD1,297 / 845+0.38 (z +0.3)+1.20 (z +0.8)-0.081R-0.036R
GBPUSD689 / 706+1.90 (z +0.9)+0.61 (z +0.3)-0.048R-0.097R
USDJPY356 / 589-3.95 (z -1.5)+6.29 (z +3.1)-0.182R+0.027R
AUDUSD762 / 538+1.46 (z +0.6)+1.97 (z +0.8)-0.043R-0.032R
XAUUSD243 / 1,632+1.32 (z +0.4)+1.23 (z +1.0)-0.073R-0.048R
BTCUSD42 / 2,075+20.24 (z +2.5)-0.30 (z -0.2)+0.221R-0.080R
US100872 / 7,468+1.28 (z +0.6)+0.71 (z +1.2)-0.074R-0.052R

iFVG, all timeframes pooled, limit entry at the near edge, 1R. BTCUSD has too few out-of-sample trades to be informative (data starts in May 2017).

No instrument shows a stable edge. USDJPY looks like a candidate in 2019–2026 (+6.29 pp, z +3.1, net +0.027R), but in 2012–2018 it is -3.95 pp (z -1.5, net -0.182R). With seven instruments in two blocks, a single strong cell that reverses in the other block is what chance alone produces.

4. Checks and corrections

  1. Independent re-simulation. A sample of 60 trades per series (EURUSD H4, USDJPY H1, XAUUSD H4) was re-run with a separate, plain-numpy implementation of the gap detection, the first-return fill and the 1R outcome on M1 data. The event counts and every sampled fill minute and outcome matched the main engine exactly (0 mismatches in 180 trades).
  2. Neutral baseline. The random entries average close to 0R before costs (-0.005R in 2019–2026, -0.009R in 2012–2018 at 1R), as they should.
  3. Consistency with the Order Blocks study. The plain FVG first return gives +1.3 pp at 1R in both blocks, the same as the 4-hour FVG result in that study, obtained here with a separate implementation and with M15, H1 and H4 pooled.
  4. A trap avoided: conditioning on the future. Splitting plain FVGs by whether they later invert produces a spectacular result: gaps that never invert win roughly 17–24 pp more than random, and those that do lose roughly 16–22 pp. That is an artefact of choosing trades by their outcome (a gap that is respected cannot also be closed through), not an edge, and it is not used anywhere in this study. The iFVG trade is entered only after the inverting candle has closed.
  5. No changes after the out-of-sample run. Definitions, bins and parameters are those fixed before the first run. The outer-quartile follow-up in Section 3.5 was added afterwards and is marked as exploratory.

5. Conclusion

  • ① The first return to an inverse fair value gap beats random entries by about one percentage point (+0.92 pp over both blocks at 1R, z +2.2). That is borderline when the blocks are pooled and not significant in either block alone, and in any case far too small to cover prop-firm costs: a one-point win-rate edge is worth about 0.02R, against costs of 0.06–0.07R. Net expectancy is negative in every timeframe and block (about -0.058R per trade pooled).
  • ② The inversion adds nothing over the ordinary gap. The plain FVG first return has the same small edge, and the difference between the two is not significant in either block.
  • ③ Entering at the middle of the gap or widening the stop does not produce an edge that holds in both blocks.
  • ④ A more forceful inversion, a larger gap or a fresher gap does not improve the result in a way that is consistent across blocks.
  • ⑤ Setups in the outer quartile of the recent range do better, in both blocks, but the same is true of ordinary gaps there, so it is an effect of location, not of inversion. It is still negative or flat after costs.

The measurable part of the inverse fair value gap is the same as for the fair value gap itself: price tends to react a little at recently formed gaps, and a little more when they sit at the edge of the recent range. The failure of the gap — the feature that defines an iFVG — did not add anything that this data could detect.

6. Limitations

  • Each element was tested with one mechanical definition (minimum gap of 0.1 ATR, inversion within 24 candles, first return within 48 candles). Discretionary traders apply the concept with context (higher-timeframe bias, session timing, nearby liquidity) that is not modelled here.
  • Events overlap in time (many gaps form and fail in the same move), so the event-level z-scores are somewhat optimistic; the block-bootstrap standard errors used throughout are the more conservative figures.
  • The minimum stop rule removes most M15 setups and many in 2012–2018; the results describe the larger gaps only.
  • The random baseline does not control for the position of the entry within the recent range. That is why the location effect in Section 3.5 appears as an edge for any gap, and why it cannot be assigned to the gap itself.
  • The random baseline uses the same UTC minute of day, so session timing shifts by one hour across daylight-saving changes. Results are on bid prices only; costs are representative estimates and slippage is not modelled. All data are from Dukascopy and differ from any broker’s or prop firm’s actual feed and costs.
  • BTCUSD covers May 2017 onwards only.

7. Next Steps

  • Build a baseline that matches the position in the recent range, to separate the location effect from the gap itself
  • Test the larger-gap subset on lower-cost venues (e.g. index futures), where the small win-rate edge might exceed costs
  • Re-test USDJPY on its own, with the two blocks treated as separate trials

Appendix — Technical notes: Gap detection runs on M15, H1 and NY-aligned H4 candles; fills and outcomes are JIT-compiled with Numba and resolved on M1 data for exact touch, stop and target ordering. ATR is the 14-candle simple average of true range on the same timeframe. Premium/discount position: middle of the gap relative to the high–low of the previous 96 candles, measured at the inverting candle (for the plain FVG, at the completion candle). Random baselines: ten same-direction, same-minute-of-day entries within ±60 days per trade, same stop distance. Standard errors: 30-day-block bootstrap. Costs: EURUSD 1.4 pips, GBPUSD 2.2 pips, USDJPY 2.5 pips, AUDUSD 1.6 pips, XAUUSD $0.58, BTCUSD $30, US100 1.5 points round trip. Full methodology and scripts available on request.