Averaging Grid Bot

The story in one line

Run exactly as specified over 3.4 years of real M15 data, the basket-grid bot does not hit its 5–7%/mo target — it loses money and blows the account. From a $5,000 start it bled to $490 (−90.20%) and hit the ruin floor on 2024-04-02, ~15 months in, then sat flat (dead). Across 1,276 baskets the win rate was 37.5%, profit factor 0.514, monthly Sharpe −5.56, max drawdown −90.26%, and 0 of 16 months were profitable (average month −13.1% against a +5–7% target).

The kill-switches, recovery throttle, two-strike override, and regime veto all fired as designed. They slowed the bleed; none of them made the bot profitable. The safety net catches the fall — it does not turn a losing strategy into a winning one.

Why it loses — the zero-cost control

Same engine, same data, peeling back one cost layer at a time to locate the loss:

Variant	Final $	Baskets	Win rate	Mean P/L / basket	Profit factor
As specified (grid + real costs)	$490	1,276	38%	$-3.53	0.51
Grid, zero cost (no spread/slip/swap)	$5,282	3,826	54%	$0.07	1.01
No grid, zero cost (pure entry edge)	$5,384	3,938	51%	$0.10	1.01

The read-out: strip the costs and the system sits at profit factor ~1.01 / ~54% win — a coin flip with no directional edge. Removing the martingale grid on top changes almost nothing (no-grid ≈ grid). The entire live loss is transaction cost applied to a zero-edge bet, compounded over thousands of baskets at roughly $3–5/basket. The spec fully defines the risk machinery but never defines an entry edge — and without one, no amount of basket, grid, or kill logic is profitable. Year by year the zero-cost edge stays flat (avg basket ≈ $0, win 52–56% in 2023–2025); there is no regime in which it earns its keep.

Backtest-vs-live: every bias points to optimism

Basket grids usually flatter themselves in backtest because historical data always retraces eventually — the simulator lets every adverse excursion mean-revert, while a live account can margin-call or stop out before the retrace arrives. Survivorship (no CHF-style de-pegs), slippage (fixed pips vs stops filling 10–50 pips worse in the exact news spikes where grids must exit), spread regime (constant vs widening 3–10× when the grid martingales hardest), and broker stop-out (a hard 10% ruin floor vs a real margin call that liquidates all seven grids at the worst tick) all point the same way: toward optimism. And the trap is that this backtest already loses — it doesn’t even get the flattering phase, because the 4h timeout and hard stops force the grid to realise its losers before the retrace. A live account would do everything this run does, only worse, with a real margin cliff the ruin floor only approximates.

Verdict: FAIL (PF 0.514). Do not deploy. Not close. The account hit the ruin floor in ~15 months, lost ~90% of capital, and was negative every month. The decisive control — strip costs and the basket sits at PF 1.01, a coin flip — shows the failure is structural: there is no entry edge for the grid to compound, only the absence of one. No risk-limit tuning changes that; the problem is upstream of risk. Basket-grid is retired.