Pipsgrowth EX10003 Momentum-Scalper

Pipsgrowth EX10003 Momentum-Scalper brands itself as an 'LSTM-risk' model, but the source is honest about a constraint: a real long short term memory network cannot be inlined into pure MQL5. What you actually get is a 5-vote momentum confluence that mimics what an LSTM directional classifier would output, namely a +1, -1, or 0 call on the last closed bar packaged with a 0-100 confidence score. For a trader evaluating the EA, that distinction matters because the bot is a momentum scalper, not a deep learning model, but the practical effect is the same: a single position is opened only when at least four of five independent momentum signals agree on direction, and that direction must also match the higher timeframe trend.

The five voters are evaluated on the bar at shift=1, so no current bar repaint is possible. The first voter is the EMA stack: EMA(8) on the signal timeframe above EMA(21) votes long, below votes short. The second is a momentum burst check: a bar whose body is larger than 0.4 * ATR(14) counts as a burst, and a bullish burst adds a buy vote while a bearish burst adds a sell vote. The third is MACD(12,26,9) histogram sign. The fourth is RSI(14) above 50 for buys and below 50 for sells. The fifth is an inline Hull MA(21) slope, computed by hand from WMA primitives inside the HullMA() function rather than via iMA, so the smoothing kernel does not depend on the indicator being available. The Hull calculation returns false on bad data rather than silently emitting a zero.

A direction is set only when one side leads by at least two votes. A 3-2 split returns 0 and no signal fires. With the default InpVoteMin of 4 out of 5, the confidence floor is 80 percent, which means a 4-1 vote clears the bar but a 3-2 split never produces a direction at all. The practical effect is that EX10003 trades only on unusually strong bar-1 agreement. Most bars do not produce a trade, which is a deliberate conservatism rather than a bug.

The confirmation layer then checks two things. First, the higher timeframe EMA(50) must agree with the direction, meaning the close at shift=1 must be above the HTF EMA for a long and below for a short. The default InpHtfTimeframe of 8 falls through the case statement to PERIOD_H1, and InpSignalTf of 0 also falls through to H1, so out of the box both default to H1. On a gold M5 chart, that means the user is effectively running a 1-hour signal with a 1-hour trend filter, not a 5-minute signal with a 1-hour filter. Changing InpSignalTf to 2 (M5) is the realistic way to use the EA on the M5 chart it is named for. Second, the implied reward to risk must clear InpMinRR of 1.2, which acts as a sanity check on the SL/TP math rather than a strategy filter.

Risk and sizing are tight by design. SL is placed at 1.8 * ATR(14) from entry and TP at 3.0 * ATR(14), so the default R:R is approximately 1.67. Lot size is fixed fractional at 0.5 percent of effective capital with a 0.01 fixed lot fallback if the risk calculation yields zero. Effective capital is the smaller of account equity and InpCapAmount (which defaults to 0 and disables the cap), plus the day's realized PnL for this magic. Daily loss limit is 3 percent and weekly is 6 percent, both computed against the same effective capital number. There is no martingale, no grid, and no pyramid; PYRAMID_MAX and PYRAMID_ATR_STEP are defined as constants but never referenced in the live code.

The regime classifier ClassifyRegime() returns one of seven states, namely StrongTrend, WeakTrend, Range, Breakout, Compress, Expand, and Choppy, based on ADX(14) at the 20 threshold, a 100-bar ATR-percentile, and Bollinger Band width. Despite being advertised as a 12-layer architecture in the header, the regime class does not actually filter entries in EX10003. The only regime checks in the entry path are a volatility floor (InpAtrPctMin of 0.10) and a hard reject on Choppy. The classifier does, however, force a position close if it transitions to Choppy on a live trade.

In-trade management runs on every tick. Three features are enabled by default. Break even moves SL to 5 points past entry once the trade reaches +1R. A 50 percent partial close fires at +1.5R, but only if the remaining volume would still be at least 1.5 times the broker's minimum lot, a small detail that prevents the partial from being rejected on micro accounts. An ATR(14) trail at 2.5x ATR, with a 100-point ratchet step, tightens SL behind price as the trade moves. Beyond those, an opposite signal close fires if the 5-vote momentum signal flips direction on a new bar, a 240-bar (about 20 hours on M5) max hold exit closes any stale position, and the Choppy regime close fires as already described.

A practical detail that matters before any live run: InpDryRun is true by default. Out of the box the EA will print what it would do and not place an order. Setting InpDryRun to false and leaving InpKillSwitch at false is required before any live trading. The InpMaxSpreadPts default of 350 is generous at roughly 35 pips on a gold symbol, which matches the volatility profile of XAUUSD on M5 and the reality that gold spread can spike into the high 20s around the New York open and the London fix. The session window of 7 to 21 in the CanTrade() gate is the London plus New York window with Asia explicitly excluded, and the weekend filter blocks Sunday and Saturday.

The source has a few non functional curiosities worth flagging. The MapTimeframeInt function is defined seven times in the file, identically, a sloppy copy paste that compiles and runs correctly but adds visual noise. NEWS_BLOCK_MINS is defined as 30 but the constant is never read anywhere in the file, so the EA has no news filter despite the constant's name. The MapTimeframeInt(0) and MapTimeframeInt(8) fall through to H1 is an additional quirk: the user must explicitly map the inputs to a value in 1 to 7 or the EA silently runs H1 logic on an M5 chart.

For a backtest, expect a moderate trade count, a low to moderate win rate, and most winners coming from the partial plus trail combination rather than from a full TP hit. The 80 percent confidence floor on entries means trade frequency will be lower than a typical scalper. The HTF EMA filter eliminates most counter trend attempts. The 3 percent daily loss cap halts the EA early on bad days, which is good for capital preservation but means the equity curve will have visible dead days. The Strategy Tester fitness function OnTester returns net*pf/(1+dd) with a 30-trade minimum, so the optimizer rewards high profit factor and penalizes drawdown rather than rewarding raw net profit alone.

In live use, a sub-100-millisecond ECN broker with raw spread on XAUUSD is the realistic minimum. Below $1,000 the partial+trail logic will frequently round the remaining lot to the broker minimum and the BE step of 5 points will often be eaten by spread on a 5-digit gold quote, so the EA tends to work best on accounts that can support a 0.10 lot without exceeding the 0.5 percent risk. Above the 0.5 percent risk, the ATR-driven stops will already be wide enough to absorb a single bad tick during NY open volatility, which is the most common loss scenario in the backtest.