When Does a Grid Basket Run On? Size, Speed and Price State in a Gold (XAUUSD) Grid-Trading Model: Backtest Evidence and a Pre-Registered Live Test

1. Introduction

A grid, or averaging-down, trading model opens an order and, if price moves against it, opens further orders on the same side at set distances, each larger than the last. All orders of the group are closed together when the group as a whole reaches a profit target. We call such a group a basket. The design makes small, frequent gains likely and large losses rare but severe, a profile it shares with martingale betting systems; Taranto and Khan (2020) relate grid-constrained strategies to the gambler’s ruin problem. For anyone who runs or monitors such a model, a practical question is how the risk still ahead of an open basket changes as the basket grows.

We study one automated gold grid model and ask three questions. Size: once a basket’s largest order has reached x lots, how often does the basket run on, and how much floating loss is still ahead of it? Speed: does the time the basket took to reach x lots carry further information? Price state: do simple measures at that moment carry further information?

Backtests of trading rules are prone to overfitting when many variants are tried (Bailey et al., 2017; Harvey, Liu and Zhu, 2016). We examined 24 candidate measures, so every backtest result here is exploratory. Confirmatory evidence will come from a live test of the running model, whose hypotheses, outcomes and decision rules were published before the first live case (Nosek et al., 2018). Live results will be reported in a later version whatever their direction.

The study measures association, not cause, and it concerns one proprietary model; the results should not be read as applying to grid strategies in general. It does not test whether acting on a signal, for example by stopping new orders, would change results.

2. Data

Model and account. The model is a proprietary gold grid model operated by the author’s company on an MT4 demo account (virtual money) with a balance of about 27 million US dollars, so 0.05% of the balance is about 13,500 dollars. Its trading rules are not disclosed; the analysis uses only its order records. Losses are reported as a share of the balance.

Orders and baskets. Every closed order from 2 January to 9 October 2026 (open and close time, side, lot size, open and close price), in broker server time (UTC+3). A basket is a set of same-side orders that close together (close times within 120 seconds). Within a basket, orders are placed in time by their open times; a basket counts as rebuilt only if every order could be placed in the one-minute price series.

Prices. One-minute XAUUSD bars (high, low) from a second broker, aligned to broker server time. Spreads, commissions and swaps are not included in the loss measure.

Two settings. On 12 August 2026 the model changed the way its order sizes step up. Before the change, adjacent lot levels were often reached by the same order, so rates per level are not comparable across settings. A basket belongs to the setting in force when its first order opened, with the split at 18:00 server time on 12 August. The account records show the new steps from 18:27, after the last basket under the old steps had closed (18:26:45), so no basket spans the change. Two old-step baskets opened between 18:00 and 18:27; one of them reached 0.15 and 0.2 lots and is therefore counted in the current setting. With 788 current-setting baskets its effect is small, but it is a known misclassification. Old setting (2 January to 12 August): 4,321 cases from 1,142 baskets. Current setting (12 August to 9 October, eight weeks): 2,091 cases from 788 rebuilt baskets that reached at least 0.15 lots (six more could not be rebuilt; see below). Only the current setting is used for the live test.

Rebuild quality. The share of baskets rebuilt was 60.3%, 60.6% and 63.4% in January, February and March and 99.5–99.7% in each month from April to September; baskets missed in January to March were larger than average, so old-setting results may be biased by selection. In the current setting six baskets that reached 0.15 lots before 9 October could not be rebuilt, four of them among the 59 baskets that reached 0.7 lots. They are excluded; Section 4.5 gives an approximate worst-case check. The order history runs twelve days past the price data and the longest closed basket lasted 73 hours, so no basket in the window was excluded for being still open.

3. Method

Cases. Levels x are 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 0.7, 1.0 and 1.5 lots, where x is the size of the largest single order so far. A basket gives one case at each level it reaches. At a single level, cases and baskets coincide; groups of levels (for example 0.15–0.25 lots) can hold several cases from one basket.

Outcomes. Run on: after the case time, the basket opens an order of at least 2x lots. Loss ahead: for each one-minute bar from the case time to the basket’s close, the floating profit or loss of the orders open in that bar is computed as the sum over orders of side × (price − open price) × lots × 100 ounces, using the bar’s low for buy baskets and high for sell baskets; the minimum of this series is the worst floating profit or loss, a negative number when the basket is under water. We report its magnitude as the loss ahead: loss ahead (% of balance) = 100 × max(0, −minimum) ÷ balance. Because the bar’s high or low is applied to all open orders at once, this worst-in-bar approximation tends to overstate the floating loss relative to prices that occurred together; it ignores trading costs and is not a bound on losses that could actually have been realised. A tail event is a loss ahead greater than 0.05% or greater than 0.2% of the balance. The tail mean (“ES 5%”) at a level is the mean loss ahead of the 5% of baskets with the largest loss ahead at that level (all baskets at the level, number of baskets in the tail rounded to the nearest whole number, at least one). The 2x rule and the two thresholds are our choices.

Speed. Minutes from the basket’s first order to its first order of at least x lots, split into thirds (fast, middle, slow) using cut-offs set within each setting and level.

Price state. 22 measures at the case time (listed in the Appendix A), including 15-minute RSI against the basket’s side, the position of price in the day’s high-low range and the size of the last 15 minutes’ move, plus the log lot level and the side: 24 inputs in all.

Estimands and inference. The size contrast is the pooled run-on share of all cases at 0.7 lots and above minus the pooled share of all cases at 0.15–0.25 lots. Speed contrasts are the pooled share of slow cases minus that of fast cases. 95% intervals are percentile intervals from 2,000 resamples of whole baskets, each resampled basket bringing all its cases (a cluster bootstrap; Efron and Tibshirani, 1993; Cameron, Gelbach and Miller, 2008). As a check, we also resampled whole calendar days of the case time. Single-level shares in Table 1 use Wilson (1927) intervals; at a single level each basket gives one case.

Combined model. A logistic regression with an L2 penalty (inverse strength C = 0.3; inputs standardised on the training data; five skewed inputs log-transformed) is fitted on the earliest 60% of cases (before 30 July 2026) and evaluated by AUC on cases from 30 July to 9 October 2026. For comparison, the same model was refitted 200 times on randomly permuted training outcomes. Because the permutation moves single cases, it does not preserve basket or day dependence, and it is reported as a descriptive benchmark, not as a test.

4. Backtest Results (Current Setting Unless Stated)

4.1. Size

Of 1,530 cases at 0.15–0.25 lots (from at most 788 baskets, since one basket can give a case at each of the three levels), 29.4% ran on; of 111 cases (from 55 baskets) at 0.7 lots and above, 48.6% did. The difference is 19.2 percentage points (95% interval 6 to 31 by baskets; 8.2 to 30.7 by days). Table 1 gives every level. The rate rises from 27.2% at 0.15 lots to about 47% at 0.7 and 1.0 lots, with wide intervals at the deep levels; we do not test for a monotonic trend. This is a conditional rate: only baskets that have already grown reach the deep levels, so it describes the risk still ahead of a basket that has got there, not a cause of that risk.

4.2. Loss Ahead

The share of baskets whose later floating loss went beyond 0.05% of the balance rises from 3.7% at 0.15 lots to 50.9% at 0.7 lots and 90.9% at 1.5 lots; for 0.2% the shares are 1.0%, 14.5% and 36.4%. The tail mean rises from 0.41% to 3.91% and 9.08% of the balance. These are interim losses of single baskets on a very large balance. They are not the account’s drawdown, which also depends on baskets open at the same time, and they say nothing about margin on a smaller account.

4.3. Speed

In the old setting, slow cases ran on more often than fast ones: 6.5 points (95% interval 1.9 to 11.8). In the current setting the overall difference is 0.1 points (−6.5 to 6.9; by days −7.2 to 7.1); sell baskets lean the same way as before (7.7; −2.2 to 18.3) and buy baskets the other way (−6.3; −15.8 to 2.8). Neither current-setting result is clear, and the old-setting result may be affected by the lower rebuild rate in January to March. Speed may also stand in for volatility, session or trend.

4.4. Price State and the Combined Model

The combined model reached an out-of-sample AUC of 0.544; one of the 200 permuted fits reached the same value. The evaluation window spans both settings, and the permutation ignores dependence, so this result is weak and exploratory; a setting-specific evaluation is left to the final data. At 1.0 lots and above, price at the day’s extreme against the basket went with more run-ons in both settings, on small samples (current setting: 61% against 39%, 28 cases each).

4.5. Robustness

Resampling whole days instead of whole baskets gives the same conclusions (Section 4.1 and Section 4.3). For the six current-setting baskets that could not be rebuilt, their order sizes are known, but their orders could not be placed in the price series. We added a case for every level each of them reached (four of them reached 0.7 lots) and set the outcomes in the worst way for the size result: all their deep cases turned back and all their shallow cases ran on. The difference then falls to 14.5 points, with a 95% interval of 2.7 to 26.3 from the same basket bootstrap on the enlarged sample. The per-level case counts of the six baskets were not kept in the analysis output for this version, so this is an approximate sensitivity check that readers cannot fully reproduce from the paper, and its interval does not include any further uncertainty about these baskets. A replay of 25 August to 9 October 2026 through the live monitoring code matched the backtest’s speed class and outcome on all 1,664 cases found by both.

5. Pre-Registered Live Test

The hypotheses below were fixed and published on the study page on 21 October 2026, before any live basket reached 0.15 lots, and registered on OSF Registries on 26 October 2026 (https://osf.io/hkvbs/; see Declarations). They are tested on the running model without any change to it. Cases are resampled by basket.

	Hypothesis	Backtest (current setting)	Confirmed if
H1 (main)	Run-on share at 0.7 lots and above exceeds that at 0.15–0.25 lots	48.6% vs 29.4% (+19.2 points)	difference at least 10 points and 95% interval above 0
H2	Slow sell cases run on more than fast sell cases	+7.7 points	difference at least 8 points and 95% interval above 0
H3	No speed difference for buy cases (H3a) or for all cases (H3b)	−6.3 / +0.1 points	each 95% interval inside −8 to +8 points
H4 (exploratory)	Tail-event shares rise with size; price at the day’s extreme at 1.0 lots and above	Table 1; Section 4.4	Benjamini–Hochberg (1995) control at 5%

Clarifications added on 26 October 2026, before the first look: H3 is decided separately for H3a and H3b. The H4 family is three tests: a rise in the share of loss ahead beyond 0.05% across levels, the same beyond 0.2%, and the day’s-extreme comparison at 1.0 lots and above.

• Sample Size and Power

At the backtest rate the account gives about 13 cases a week at 0.7 lots and above, from about 7 baskets, and about 185 cases a week at 0.15–0.25 lots, from about 95 baskets. Clustering inflates the variance of the shares by a factor (design effect) of 1.6 at the deep levels and 2.1 at the shallow levels. The final read is after 15 weeks, in early February 2027. Power was simulated by drawing baskets with replacement from the current-setting backtest at the weekly rates above for the stated number of weeks, 1,000 times, and computing in each simulated sample a 200-resample basket-bootstrap interval for the difference. The share of simulated samples whose lower bound was above 0 was 0.95 at 15 weeks if the true difference equals the backtest’s, and 0.35 at 15 weeks (0.48 at 20 weeks) if it is 10 points. These figures cover only the “interval above 0” part of the H1 rule; power for the full rule (which also needs a difference of at least 10 points) was not computed and can be no higher. They also assume the backtest’s basket distribution. A hypothesis that misses its rule will be reported as not confirmed, not as evidence of no effect. An interim look on 18 November 2026 is descriptive only, with no test and no stopping.

• Live Data

The account reports its open orders every minute to a monitoring server, which rebuilds baskets and records each level time, speed class, price measures and floating loss. Live price measures come from the open orders themselves and are taken only when a report arrives within three minutes of the order, so live measures are close to, but not the same as, the backtest’s. Baskets that open and close within one minute are not seen; they rarely reach 0.15 lots. Open baskets at the final read will be kept as unfinished cases.

6. Limitations

One model, one broker and a demo account, over ten months that included unusually large gold moves; the current setting has eight weeks of history. Prices come from a second broker and the loss measure ignores spreads, commissions and swaps. Rebuild quality was lower in January to March. The run-on rule and the loss thresholds are choices, and other choices could give different rates. The analysis is associational; it does not show results on a smaller account, where margin limits could end a basket earlier.

7. Conclusions

In this model’s backtest, a basket that has already grown further was more likely to keep growing, and its remaining loss rose faster than its size. Of the measures examined, the level already reached showed the strongest association with the risk still ahead; speed of growth and the price-state measures added little in the current setting. Whether the size association holds on new data is the subject of the pre-registered live test, whose results will be reported in full.