forked from animatedread/Warrior_EA
10253b581b
User request: attaching a chart must need zero Inputs-tab edits. Private (non-Market) build now defaults to AIType=META, all four classic families ON (they are the sweep's candidate sources), Meta_ExportDataset=true. Market-build defaults unchanged (AI_NONE, MA/RSI only, no export); UseDatabaseRanking=true applies to both per the earlier request. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
682 lines
59 KiB
MQL5
682 lines
59 KiB
MQL5
//+------------------------------------------------------------------+
|
|
//| Inputs.mqh |
|
|
//| AnimateDread |
|
|
//| https://www.mql5.com |
|
|
//+------------------------------------------------------------------+
|
|
#property copyright "AnimateDread"
|
|
#property link "https://www.mql5.com"
|
|
#include "..\Enumerations\InputEnums.mqh"
|
|
//--- Each `input string *_Settings` below is a GUI-only section divider: MetaTrader renders an input
|
|
//--- string whose value equals its comment as a header. Never read by MQL5 code - that's expected, not
|
|
//--- dead wiring. Sections are ordered most-used first: General, Money, Trade, Classic Signals,
|
|
//--- Neural Network, AI Input Features, Filters, Trade Journal - then NN Optimizer / Performance LAST.
|
|
//--- The Neural Network block sits directly ABOVE AI Input Features because that is the reading order a
|
|
//--- user actually needs: choose the architecture, then choose what it sees. NN Optimizer / Performance
|
|
//--- must remain the final divider in this file - the Adam/Sgd inputs are declared in AI\Network.mqh and
|
|
//--- render immediately after it, so anything added below would land inside that group.
|
|
//==================================================================================================
|
|
// GENERAL
|
|
//==================================================================================================
|
|
input string Expert_Settings = "General"; // General
|
|
input ulong Expert_MagicNumber = 2024; // Magic number (unique EA id)
|
|
input bool Expert_EveryTick = false; // Calculate on every tick
|
|
//--- NO LONGER AN INPUT (2026-08-01). The detailed panel/journal is a developer view: a buyer does not
|
|
//--- care which plateau stage the ladder is on, and every row in the Inputs tab is a row they have to
|
|
//--- read past to reach something that matters. Same reasoning that already applied to DebuggingMode
|
|
//--- below, so the two now sit together. Flip to true and recompile to work on the EA.
|
|
const bool VerboseMode = false;
|
|
//--- DELIBERATELY NOT AN INPUT. Development diagnostics: dumps the training internals that used to sit on
|
|
//--- the on-chart panel (plateau-ladder stage, eras-since-best, the deploy gate, selection internals) into
|
|
//--- the Experts journal instead, where they cost the user nothing. This is a commercial product - the
|
|
//--- default panel has to read like a product, not like a training console, so anything a buyer cannot act
|
|
//--- on belongs in a log. Flip to true and recompile when diagnosing a training run.
|
|
const bool DebuggingMode = false;
|
|
//--- ALSO NOT AN INPUT, and for a stronger reason than DebuggingMode. Pins the dense-taper depth instead
|
|
//--- of deriving it (ComputeHiddenLayerCount), purely so a depth comparison can still be run while
|
|
//--- working on the EA. 0 = derived, which is the only value that should ever ship. A user who picks a
|
|
//--- depth is contradicting the first-layer width and the taper the code derived around it - that
|
|
//--- contradiction is exactly what the MLP_3L/MLP_4L presets used to allow.
|
|
//--- NOTE the limitation: this is compile-time, and it feeds the weights-filename fingerprint only when
|
|
//--- non-zero, so two forced depths get their own model files but cannot run SIMULTANEOUSLY from one
|
|
//--- .ex5. Depth comparisons are sequential unless you deploy two separately-compiled builds.
|
|
const int ForceHiddenLayers = 0;
|
|
//==================================================================================================
|
|
// MONEY MANAGEMENT
|
|
//==================================================================================================
|
|
input string MM_Settings = "Money Management"; // Money Management
|
|
input MONEY_MANAGEMENT_STRATEGY MM_STRATEGY = FIXED_RISK; // MM strategy
|
|
input MONEY_RISK_PERCENT_PRESET Money_Risk_Percent = RISK_PCT_1; // Risk % of balance per trade
|
|
input double Money_FixLot_Lots = 0.01; // Fixed lot size [0.01-10]
|
|
//==================================================================================================
|
|
// TRADE MANAGEMENT (entry / stop / target / trailing / exit)
|
|
//==================================================================================================
|
|
input string Entry_Settings = "Trade Management"; // Trade Management
|
|
input TRADING_DIRECTION tradingdirection = BOTH; // Trade direction
|
|
//--- ENTRY / STOP / TARGET ARE NO LONGER INPUTS (2026-08-07). They were three enums the user had to pick,
|
|
//--- and in the tester they were three more axes for a genetic optimization to overfit. The barrier
|
|
//--- geometry is now MEASURED (ReportBarrierGeometryScan picks the SL:TP pairing that carries the most
|
|
//--- entry-time information about its own outcome, and only adopts it when it clears a family-wise
|
|
//--- significance gate - otherwise these defaults stand). Kept as named constants rather than deleted so
|
|
//--- every existing reference still reads the same, and so the fallback is stated in one place.
|
|
//---
|
|
//--- Entry is pinned to MARKET deliberately. The pending-order modes place the entry at a LEVEL while the
|
|
//--- rest of the pipeline measures from the bar open, which is precisely the mismatch that manufactured
|
|
//--- the +0.097 R "retail fade" result later retracted as a fill artifact - a pending entry cannot be
|
|
//--- honestly simulated by this codebase's own fill model, so it is not offered.
|
|
const ENTRY_MULTIPLIER Entry_Multiplier = MARKET; // Entry type/offset (fixed - see above)
|
|
//--- STARTING geometry only. The scan may replace this pair at era 0 on a fresh model; a model that has
|
|
//--- already been trained reads its pinned pair back out of the .cfg and never re-measures, so the labels
|
|
//--- a run started with are the labels it finishes with.
|
|
//---
|
|
//--- MIN REWARD:RISK IS GONE (2026-08-09). It was the last place a GUESS could override a MEASUREMENT.
|
|
//--- The barrier geometry is derived from the instrument's own excursion distribution - stop at q75 of
|
|
//--- adverse travel, target at q50 of favourable - and then a 1:2 floor was applied on top, raising the
|
|
//--- target to whatever twice the stop happened to be. On SP500 H1 that turned a reachable target into
|
|
//--- 6.66*ATR, which only 3.3% of bars reach inside the horizon: the label became "almost never a win",
|
|
//--- and the model was trained to predict an event that essentially does not occur.
|
|
//---
|
|
//--- The ratio never bought anything it was believed to buy. A reward:risk floor does not create
|
|
//--- expectancy - it trades hit rate for payoff at a fixed break-even (see the barrier-geometry log
|
|
//--- line, which prints that break-even next to the ranking precisely to make this visible), and this
|
|
//--- project has already MEASURED that exit shape moves payoff without moving expectancy at all. What
|
|
//--- it did buy was two outages: four consecutive Market validation rejections for "no trading
|
|
//--- operations" when it rejected 100% of setups, and the label corruption above.
|
|
//---
|
|
//--- Risk is controlled where risk is actually controlled - the per-trade account risk percentage and
|
|
//--- CRiskBudget's daily/total drawdown enforcement - not by a ratio filter at the door.
|
|
const STOP_LOSS_MODE SL_Mode = SL_ATR_x2; // Stop-loss mode (measured - see above)
|
|
const TAKE_PROFIT_MODE TP_Mode = TP_ATR_x6; // Take-profit mode (measured - see above)
|
|
input TRAILING_STRATEGY TrailingStrategy = TRAILING_STRATEGY_NONE; // Trailing stop
|
|
input BARS_EXPIRATION Signal_Expiration = BARS_X3; // Pending order expiry (bars)
|
|
input CONFIDENCE_SOURCE Confidence_Source = CONF_AI; // AI confidence source (SL/TP/trail/exit/MM)
|
|
//--- UNIFIED conviction gates - ONE pair of thresholds governing BOTH engines, classic and AI. There
|
|
//--- used to be a second, AI-only pair in the Neural Network section (Min AI confidence / Min AI exit
|
|
//--- confidence) duplicating these: four inputs for what is really two decisions, where a trader could
|
|
//--- set the vote gate and still be silently overruled by the AI floor (or the reverse). Merged here.
|
|
//--- Everything is expressed on the same 0-100 conviction scale: a classic filter contributes its
|
|
//--- pattern weight (10-100), an AI signal contributes its confidence tier (80-100), and
|
|
//--- CExpertSignalCustom::Direction() averages the filters that voted before
|
|
//--- CheckOpenPosition/CheckClosePosition threshold that average.
|
|
//--- Open - aggregate conviction required to ENTER, and NOTHING else. It has exactly one meaning for
|
|
//--- both engines: the averaged vote across the filters that voted must reach it.
|
|
//--- It used to do two further jobs on the AI side - an entry floor on the winning softmax
|
|
//--- probability, and the base the 4 AI confidence tiers were quartiled from - which put one
|
|
//--- number on two incompatible scales. A 3-class argmax winner is arithmetically >= 1/3, so
|
|
//--- as a floor every setting from 0 to 33 gated precisely nothing, while every setting above
|
|
//--- that ALSO silently moved the tier boundaries. Both jobs are gone. The AI now expresses
|
|
//--- confidence the way a classic signal does - as the WEIGHT of the vote it casts, 25/50/75/
|
|
//--- 100 across its four tiers, quartiled from the head's own structural floor (1/3 for the
|
|
//--- 3-class softmax, 0.5 for the regression head - see CExpertSignalAIBase::ConfidenceTier).
|
|
//--- So this input now reads, for the AI voting alone: 25 = trade any directional call,
|
|
//--- 50 = tier 1 and up, 75 = tier 2 and up, 100 = only near-certain calls. A weak AI call is
|
|
//--- no longer blocked inside the AI - it votes weakly and is filtered here, exactly like a
|
|
//--- weight-10 classic confirmation.
|
|
//--- NOTE in a hybrid setup this is an AVERAGE: a tier-3 AI vote of 100 alongside two
|
|
//--- weight-10 classic confirmations averages to 40, not 100. Raising this input while several
|
|
//--- low-weight classic signals are enabled suppresses strong AI calls by dilution - that is
|
|
//--- inherent to averaging, and it is the same arithmetic the classic-only path has always had.
|
|
//--- Close - OPPOSITE conviction required to EXIT. It drives BOTH exit routes, at the same conviction:
|
|
//--- the averaged rule-based vote, and the AI early exit (how strongly the AI must have flipped
|
|
//--- AGAINST an open position before that alone closes it). There is deliberately no separate
|
|
//--- "Early AI exit" switch any more - it was a third input for what these two routes already
|
|
//--- express, and it could be left off while Close was set, silently discarding the exit the
|
|
//--- trader had just asked for. The two routes are NOT redundant with each other and both are
|
|
//--- needed: the AI's normal vote is one-shot (LongCondition/ShortCondition consume the
|
|
//--- m_lastNonNeutralSignal alternation gate when they fire) and is then AVERAGED with every
|
|
//--- other filter, so an AI reversal that gets diluted below Close on the bar it happens is
|
|
//--- consumed and never re-offered, leaving the position open indefinitely. The early-exit
|
|
//--- route reads the AI's LIVE signed confidence every bar, undiluted, and so still fires.
|
|
//--- Set Close = Disabled to switch off vote-driven exits entirely (SL/TP/trailing only) -
|
|
//--- that turns off both routes at once, since 101 is unreachable on either scale. See
|
|
//--- VOTE_CLOSE_PRESETS in Enumerations\InputEnums.mqh.
|
|
//--- Close defaults ABOVE Open deliberately: a position is an existing commitment with real cost to
|
|
//--- abandon, so reversing out of one should demand more conviction than opening it did, and a signal
|
|
//--- hovering either side of the entry gate must not be able to churn a position open and shut. Both
|
|
//--- were once hardcoded to 10/10 - one value for BOTH directions of the decision, pinned at the LOWEST
|
|
//--- weight any pattern can carry - so with MA/RSI Pattern_0 (weight 10) firing on nearly every bar on
|
|
//--- whichever side of the MA price sits, one cross flipped the average from +10 to -10 and closed the
|
|
//--- position on the very next bar. The stock MQL5 wizard makes the same asymmetric choice, 50 to open
|
|
//--- against 100 to close.
|
|
input PERCENTAGE_PRESETS Min_Vote_Open = PCT_20; // Min vote to open - AI + classic
|
|
input VOTE_CLOSE_PRESETS Min_Vote_Close = VOTE_CLOSE_DISABLED; // Min opposite vote to close - AI + classic
|
|
//==================================================================================================
|
|
// CLASSIC SIGNALS (rule-based MA/RSI votes - trade alongside or instead of the neural network)
|
|
//==================================================================================================
|
|
input string Classic_Settings = "Classic Signals"; // Classic Signals
|
|
//--- EnableMA/EnableRSI default depends on the build (see AIType's declaration comment for the full
|
|
//--- rationale): ON for a Market submission build (WARRIOR_MARKET_BUILD defined) so a fresh install
|
|
//--- trades immediately with no AI warm-up, OFF for the private/live build, which runs AI-only by
|
|
//--- default. Either way this is only a compile-time DEFAULT - still a normal input, changeable per-run
|
|
//--- from the Inputs tab without recompiling.
|
|
//--- PRIVATE-BUILD DEFAULTS CHANGED 2026-08-13 (user request, meta-pooling campaign): all four
|
|
//--- classic families default ON in the private build - they are the META chart's candidate sources
|
|
//--- for the on-chart ladder sweep (BuildCorpusBySweep), and the drop-on-chart workflow must need no
|
|
//--- Inputs-tab edits. Market build unchanged: MA/RSI on (trades out of the box), MACD/Ichimoku off
|
|
//--- (additive to the shipped strategy).
|
|
#ifdef WARRIOR_MARKET_BUILD
|
|
input bool EnableMA = true; // MA classic vote
|
|
#else
|
|
input bool EnableMA = true; // MA classic vote
|
|
#endif
|
|
#ifdef WARRIOR_MARKET_BUILD
|
|
input bool EnableRSI = true; // RSI classic vote
|
|
#else
|
|
input bool EnableRSI = true; // RSI classic vote
|
|
#endif
|
|
#ifdef WARRIOR_MARKET_BUILD
|
|
input bool EnableMACD = false; // MACD classic vote
|
|
input bool EnableIchimoku = false; // Ichimoku classic vote
|
|
#else
|
|
input bool EnableMACD = true; // MACD classic vote
|
|
input bool EnableIchimoku = true; // Ichimoku classic vote
|
|
#endif
|
|
//--- Indicator parameters below are SHARED: they define the classic votes above AND seed the matching AI
|
|
//--- input features (AI Input Features section) as their starting period, which Auto-tune indicators then
|
|
//--- searches from. Set once here, used by whichever consumer(s) are enabled.
|
|
input MA_PERIOD_PRESETS PeriodMA = MA_PERIOD_50; // MA period
|
|
input MA_TYPE_PRESETS MA_Type = MA_TYPE_SMA; // MA type (SMA/EMA/.../T3/Kalman)
|
|
input RSI_PERIOD_PRESETS PeriodRSI = RSI_PERIOD_14; // RSI period
|
|
//--- MACD AND ICHIMOKU PERIODS ARE NO LONGER INPUTS (2026-08-01). Six dropdowns, pinned here at the
|
|
//--- textbook values every reference uses (12/26/9 and 9/26/52), for three reasons:
|
|
//--- 1. They were six of the largest contributors to the Inputs tab, for indicators that both ship
|
|
//--- DISABLED. Rows a user must scroll past to reach the AI settings are a real cost.
|
|
//--- 2. As optimizer inputs they are an overfitting surface. A genetic sweep across 6 period
|
|
//--- dimensions on one symbol's history will always find a combination that looks excellent and
|
|
//--- generalizes to nothing - and it costs nothing to discover, which is what makes it dangerous.
|
|
//--- 3. They are the SEED for the AI's own auto-tuner (AutoTuneIndicators), which searches from these
|
|
//--- values against a held-out objective. That search is the supported way to move them: it is
|
|
//--- validated, it is per-model, and it cannot silently overfit the way a raw optimizer pass can.
|
|
//--- Left as named constants rather than deleted because they are still read in both roles (classic
|
|
//--- vote periods AND auto-tune starting points), and because the classic textbook values are the
|
|
//--- correct fixed answer for a vote that exists mainly to satisfy marketplace validation.
|
|
const MACD_FAST_PRESETS MACD_PeriodFast = MACD_FAST_12;
|
|
const MACD_SLOW_PRESETS MACD_PeriodSlow = MACD_SLOW_26;
|
|
const MACD_SIGNAL_PRESETS MACD_PeriodSignal = MACD_SIGNAL_9;
|
|
const ICHIMOKU_TENKAN_PRESETS Ichimoku_PeriodTenkan = ICHI_TENKAN_9;
|
|
const ICHIMOKU_KIJUN_PRESETS Ichimoku_PeriodKijun = ICHI_KIJUN_26;
|
|
const ICHIMOKU_SENKOU_PRESETS Ichimoku_PeriodSenkou = ICHI_SENKOU_52;
|
|
//==================================================================================================
|
|
// NEURAL NETWORK (training)
|
|
//==================================================================================================
|
|
input string NNetworks_Settings = "Neural Network"; // Neural Network
|
|
//--- AIType default depends on the build, via the same WARRIOR_MARKET_BUILD compile-time flag that
|
|
//--- strips the DLL import block for Market submissions (see Warrior_EA.mq5's top-of-file comment) - not
|
|
//--- an input value itself (that can't be set programmatically), only which default the Inputs tab
|
|
//--- starts on:
|
|
//--- - WARRIOR_MARKET_BUILD defined (Market submission): OFF - a fresh install trades from Classic
|
|
//--- Signals (MA/RSI) out of the box with no AI warm-up, satisfying MQL5's automated check for live
|
|
//--- trade activity within its test window.
|
|
//--- - Not defined (private/live build): MLP - this build runs AI-only from the start, with Classic
|
|
//--- Signals defaulting off too (see EnableMA/EnableRSI), so no per-run manual input changes are
|
|
//--- needed switching between preparing a submission and running the real thing.
|
|
//--- Either way, still a normal input - freely changeable per-run from the Inputs tab.
|
|
#ifdef WARRIOR_MARKET_BUILD
|
|
input AI_CHOICE AIType = AI_NONE; // AI architecture preset (or Disabled)
|
|
#else
|
|
//--- META since 2026-08-13 (was HYBRID): the private build's active campaign is the meta-labeling
|
|
//--- pool - drop a chart on any symbol and it sweeps candidates, labels, trains and exports with no
|
|
//--- Inputs-tab edits. Direction models remain selectable per-run as always.
|
|
input AI_CHOICE AIType = AI_META; // AI architecture preset (or Disabled)
|
|
#endif
|
|
//--- SGD or ADAM weight update (honored by PAI/CONV/LSTM/HYBRID). SGD rate/momentum are AI\Network.mqh inputs.
|
|
//--- A third "DFA" option was briefly the default (2026-07-28) and has been removed - it was a
|
|
//--- deterministic index-parity sign flip on the gradient, i.e. permanent gradient ASCENT on half of
|
|
//--- every weight tensor, and its backward pass was structurally incompatible with the OpenCL/DirectML
|
|
//--- neuron model. See ENUM_OPTIMIZATION's comment in AI\Network.mqh.
|
|
input ENUM_OPTIMIZATION TrainingOptimizer = ADAM; // Weight optimizer
|
|
//--- OUTPUT TYPE IS NO LONGER AN INPUT (2026-08-01). The regression head (1 tanh output) was an option
|
|
//--- that never made sense for the question this system asks. Since the triple-barrier relabel the
|
|
//--- target is explicitly an EVENT - "does a trade opened here reach its target before its stop" - and
|
|
//--- the right output for an event is its probability, which is what the 3-class softmax head produces.
|
|
//--- A regression head would have to predict a continuous quantity that the label does not even contain,
|
|
//--- and every downstream consumer already speaks probability: the confidence tiers quartile the softmax
|
|
//--- winner, dir-precision is a win rate over called bars, and the class priors calibrate a distribution.
|
|
//--- The regression path is still IMPLEMENTED throughout (m_outputNeuronsCount == 1 branches, the 0.50
|
|
//--- magnitude cutoff in DoubleToSignal) and is left in place deliberately: it costs nothing dormant and
|
|
//--- removing it would touch every scoring path at once for no gain. It is simply no longer selectable.
|
|
const OUTPUT_NEURONS_COUNT OutputNeuronsCount = OUTPUT_CLASSIFICATION;
|
|
//--- No "first layer neurons" input any more. Its only defensible value depends on two things the user
|
|
//--- cannot see - the input-vector width after feature selection, and how much in-sample data the study
|
|
//--- period yields - so it is derived at topology-build time instead. See
|
|
//--- CExpertSignalAIBase::ComputeFirstLayerWidth(). The old default (500) was ~8 parameters per training
|
|
//--- sample and expanded a 420-wide correlated input rather than compressing it.
|
|
//--- No "LSTM hidden size" or "CONV filter count" inputs either, removed 2026-07-30 for exactly the
|
|
//--- reason above: both defaulted to a fixed constant (32 units, 16 filters) chosen without reference to
|
|
//--- the input they sit on, which is the one thing that decides whether either number is sane.
|
|
//--- The conv layer is a per-bar projection (window = step = one bar's features), so 16 filters
|
|
//--- COMPRESSED a 50-feature configuration but EXPANDED a minimal 4-feature one 4x - adding parameters
|
|
//--- below every learnable layer without adding information. The LSTM block is worse: its weight count
|
|
//--- is 4*H*(H+inputs+1), so 32 units against a 540-wide input is ~73k weights, more than double the
|
|
//--- entire derived dense taper it feeds, and it was the one stage the capacity budget never covered.
|
|
//--- Both are now derived from the per-bar feature count and the same one-weight-per-training-bar budget
|
|
//--- the first layer uses. See ComputeConvFilterCount()/ComputeLstmHiddenSize().
|
|
//--- ConvPoolWindow / ConvPoolStep removed 2026-07-29. The pooling stage they configured reduced
|
|
//--- across FILTER channels rather than across time - a consequence of the conv layer's position-major
|
|
//--- output layout that no window/step pair can correct. See AddConvStage() in Expert\ExpertSignalAIBase.mqh.
|
|
//--- No "min neurons" / "reduction per layer" inputs either. With the first layer's width derived
|
|
//--- (ComputeFirstLayerWidth) the taper has no freedom left: it runs geometrically from that width down
|
|
//--- to a final hidden layer sized off the output count, spread over the layer count the chosen AIType
|
|
//--- implies. Keeping either knob would let the user contradict the derivation - and both were
|
|
//--- calibrated for the old hand-picked 500-wide first layer, where they gave 500->150->45; against the
|
|
//--- derived 64 they degenerate to 64->20->20. See BuildFreshTopology()'s taper block.
|
|
//--- Batch normalization (Ioffe & Szegedy 2015) between every pair of dense layers, including just
|
|
//--- before the classification head. ON by default: without it the only bounded stage in the whole
|
|
//--- forward path was the sigmoid head, and the observed failure mode ordered exactly by depth - the
|
|
//--- shallow perceptron held ~52% balanced accuracy while the deepest topology sat on the 33.3%
|
|
//--- one-class floor. It also decouples WEIGHT_DECAY from the learned function, which is what stops
|
|
//--- the slow monotonic decay of the per-bar logit spread that preceded every collapse.
|
|
//--- Left as an input rather than hardcoded so the effect can be A/B'd without a recompile. It is part
|
|
//--- of the weights-filename fingerprint, so flipping it starts a separate model rather than resuming
|
|
//--- an incompatible one. See AI\NeuronBatchNorm.mqh.
|
|
//--- NOT an input. Batch normalization is required, not optional: measured 2026-07-29 on identical
|
|
//--- MLP_3L topologies it was worth +11.3 points of balanced accuracy (57.0% with, 45.7% without),
|
|
//--- stable across 150+ and 200+ eras, and the no-BN control converged to ~5% IS and OOS accuracy
|
|
//--- with no chart signals at all. A user cannot make a good decision here and can easily make a
|
|
//--- ruinous one, so the choice is not offered. Kept as a named constant rather than deleted: the
|
|
//--- topology builder, the weights fingerprint and the .cfg guard all read it, and a constant keeps
|
|
//--- those paths (and the ability to flip it for a diagnostic rebuild) intact.
|
|
const bool EnableBatchNorm = true; // AI: batch normalization
|
|
//--- EMA window the running mean/variance are estimated over, in TRAINING SAMPLES (bars replayed),
|
|
//--- not eras. Training here is pure online SGD - one update per sample - so there is no mini-batch to
|
|
//--- average over and this stands in for the batch size. Long enough to be a stable estimate of the
|
|
//--- feature distribution, short enough to track a genuine regime change. 1000 is ~3% of a typical
|
|
//--- 36k-bar in-sample window.
|
|
//--- Also not an input, for the same reason plus one more: this is a running-statistics window in
|
|
//--- SAMPLES, and nothing on the Inputs tab tells a trader what a good value is. It only ever had
|
|
//--- two meaningful settings - large enough to be a stable estimate, or <=1 which silently disables
|
|
//--- the layer entirely. The first is the only correct one.
|
|
const int BatchNormWindow = 1000; // AI: batch-norm window (samples)
|
|
//--- No "training years" input either, removed 2026-07-30. There is no case for training on less data
|
|
//--- than the broker actually provides: this is a weak signal at a ~6% directional base rate, every extra
|
|
//--- year is more of the minority class, and the honest generalization read comes from the out-of-sample
|
|
//--- holdout below rather than from withholding history. Training now starts at the earliest available
|
|
//--- bar (floored by MinTrainYear, which exists to exclude a broker's dubious pre-history, not to size
|
|
//--- the run). The topology's capacity budget reads the REAL bar count that yields - see
|
|
//--- CExpertSignalAIBase::EstimatedInSampleBars(), and the note there on why that measurement is taken
|
|
//--- exactly once and then pinned.
|
|
input OOS_SPLIT_PRESET OOSSplit = OOS_30; // Out-of-sample holdout
|
|
//--- There is deliberately NO "target accuracy" input. Training runs until it stops improving and then
|
|
//--- deploys its own best model: after a stretch of eras with no new best it tries to escape the plateau
|
|
//--- (learning-rate warm restart, then focal-gamma anneal), and if neither finds anything better it
|
|
//--- finalises the best checkpoint it found. See the PLATEAU_* ladder in Expert\ExpertSignalAIBase.mqh.
|
|
//--- An absolute target could only ever be wrong in one of two directions: set above what a given
|
|
//--- symbol/timeframe can reach and the run never converges (it burns to the era cap and deploys the same
|
|
//--- checkpoint hours later anyway); set below and it stops a run that was still getting better.
|
|
//--- MinRecall stays, and is NOT a performance target - it is the anti-collapse floor that makes
|
|
//--- auto-deploy safe. Buy, Sell AND Neutral must each be recognised this well on held-back data before a
|
|
//--- checkpoint is eligible to ship, so a model that quietly gives up on one direction can never deploy.
|
|
//--- It is an OOS CLASSIFICATION metric (3-class), NOT a trade win rate: random guessing is ~33%.
|
|
//--- 2026-07-29: 60 -> 40. 60 was never demonstrated reachable on this data. The ONE successful
|
|
//--- auto-deploy in the logs (Hybrid, SP500 H1, 28th 00:50, best balanced 66.0%) ran against a 40%
|
|
//--- floor; every run since has been gated at 60 and none has come close - CONV/LSTM/Hybrid peaked at
|
|
//--- 40/49/41% balanced and then decayed, so stage 3 refused to deploy and reset the ladder ~27 times,
|
|
//--- turning a converged run into a 1000-era one-way trip. A floor above what the configuration can
|
|
//--- reach is exactly the "absolute target set too high" failure the comment above warns about, just
|
|
//--- expressed per-class. Raise it again only after a run actually clears it with headroom.
|
|
//--- 2026-08-01: NO LONGER AN INPUT. This is a safety floor, not a preference, and the one direction a
|
|
//--- user can move it is the harmful one - raising it past what the configuration reaches does not
|
|
//--- produce a better model, it produces NO model (nothing clears the gate, stage 3 refuses to deploy,
|
|
//--- and the run burns to the era cap). That failure was observed repeatedly at 60 and is the exact
|
|
//--- catch-22 this floor was nearly deleted over. 40 is the value the only successful auto-deploy in the
|
|
//--- project's history ran against. It also no longer decides what SHIPS - deployability moved to
|
|
//--- directional precision with a derived coverage floor - so it now only drives the diagnostic recall
|
|
//--- line, which makes exposing it even harder to justify.
|
|
const PERCENTAGE_PRESETS MinRecall = PCT_40;
|
|
//--- There are deliberately NO AI-only confidence inputs here any more. The AI entry floor and the AI
|
|
//--- early-exit threshold are the SAME two numbers the classic votes use - Min vote to open / Min
|
|
//--- opposite vote to close (Trade Management section) - so one pair of inputs governs both engines;
|
|
//--- see their declaration comment for how the 0-100 scale maps onto AI softmax confidence and tiers.
|
|
//==================================================================================================
|
|
// CLASS IMBALANCE - ONE MECHANISM, ONE KNOB
|
|
//==================================================================================================
|
|
//--- The directional base rate here is ~3% Buy / ~3% Sell / ~94% Neutral (a ~31:1 imbalance), so the
|
|
//--- loss needs SOME correction or the optimum is "always predict Neutral". This section used to offer
|
|
//--- NINE inputs for that one job. They were consolidated on 2026-07-31 because, audited against the
|
|
//--- code, five of them did not do what their names said at the shipped defaults:
|
|
//--- AILogitPriorStrength DEAD - Inference.mqh's post-hoc prior early-returns when the adjusted
|
|
//--- loss is on, because the offsets are already trained into the weights.
|
|
//--- OversampleParity DEAD in training - Training.mqh gates the replay loop on
|
|
//--- !useLogitAdjustedLoss (correctly: Buda et al. 2018 on why stacking
|
|
//--- oversampling with an analytic correction double-counts the imbalance).
|
|
//--- EnableMinorityReplay DEAD as replay; it survived ONLY as a focal-gamma damper (x0.125).
|
|
//--- ConstrainReplay DEAD as a cap; it only chose between damper 0.125 and 0.25.
|
|
//--- UseStaticPrior an exact duplicate of FreezePriorCalibration - the two were OR'd
|
|
//--- together in the single place either was read.
|
|
//--- Focal loss was the one real redundancy: it ran at gamma*0.125 alongside the adjusted loss, i.e.
|
|
//--- two corrections on the SAME axis, which is what the codebase's own Buda et al. citation warns
|
|
//--- against. Removed rather than re-tuned - the plateau ladder's escape is its learning-rate warm
|
|
//--- restart, and the gamma anneal it also performed was only ever a monotone step toward zero.
|
|
//---
|
|
//--- WHAT REMAINS is LOGIT-ADJUSTED LOSS (Menon et al. 2021, ICLR, "Long-tail learning via logit
|
|
//--- adjustment"): add tau*log(prior_c) to each class logit inside the TRAINING gradient only. The
|
|
//--- network learns to absorb the offset, so at inference its RAW argmax is already the
|
|
//--- balanced-error-optimal decision - no second correction at read time, by construction. Minimizing
|
|
//--- softmax cross-entropy on adjusted logits is consistent for BALANCED error, which is the metric
|
|
//--- checkpoint selection already ranks on, so the loss and the deploy decision optimize the same
|
|
//--- thing. It is the only one of the six with a consistency guarantee, which is why it is the one kept.
|
|
//---
|
|
//--- tau: 100% = tau 1.0, the paper's default and the only value carrying the guarantee. 0 = OFF, which
|
|
//--- is now the honest way to disable the correction entirely (it replaces the old EnableLogitAdjusted-
|
|
//--- Loss boolean - a separate on/off switch beside a strength dial where 0 already means off is two
|
|
//--- controls for one decision). NOTE the runtime auto-caps tau so the offsets cannot swamp the output
|
|
//--- head's usable logit range; the startup line reports the capped value actually used.
|
|
input LOGIT_PRIOR_STRENGTH_PRESETS LogitAdjustTau = LOGIT_PRIOR_100; // AI: class-imbalance correction (tau, 0=off)
|
|
//--- Stop EMA-updating the measured class priors after the first real measurement. NOT AN INPUT as of
|
|
//--- 2026-08-01: it answers a question a user has no way to evaluate ("should the correction track this
|
|
//--- era's tally or the first one's"), and after the triple-barrier relabel the priors barely move
|
|
//--- between eras anyway - the labels are near-balanced and stable, which is the whole point of the
|
|
//--- relabel. Letting them track is the correct default; freezing exists for a symbol whose class
|
|
//--- distribution genuinely shifts mid-run, which is a developer's diagnostic, not a product setting.
|
|
const bool FreezePriorCalibration = false;
|
|
//--- SWING CONFIRM BARS IS NO LONGER AN INPUT (2026-08-01), but the constant is still load-bearing and
|
|
//--- must not be deleted. It stopped gating the LABELS with the triple-barrier relabel - that lookahead
|
|
//--- is now the barrier horizon, which is measured (ComputeBarrierHorizonBars) rather than configured.
|
|
//--- It still gates the swing-context INPUT FEATURES (EnableSwingContext, on by default, 9 features):
|
|
//--- ZigZag revises its most recent legs, so a feature that read the raw current buffer would be reading
|
|
//--- a value the live bar could not actually have had yet. That is straight lookahead into the feature
|
|
//--- vector, so this embargo stays - it simply has no reason to be user-facing, because the correct
|
|
//--- value is a property of the ZigZag indicator's own recalculation depth, not of anyone's preference.
|
|
//--- Kept in the weights fingerprint at its shipped value, so pinning it re-keys nothing.
|
|
const SWING_CONFIRMATION_PRESET SwingConfirmationBars = SC_100;
|
|
//--- Continual learning: after the model is deployed, keep adapting it on a LIVE chart to newly-RESOLVED
|
|
//--- market structure - the same supervised triple-barrier task it was trained on, waiting the full
|
|
//--- barrier horizon so a bar whose outcome is not yet decided is never learned from. The deployed model
|
|
//--- only moves toward the update while a rolling-accuracy guardrail holds; if accuracy decays the blend
|
|
//--- FREEZES (live keeps trading the last-good shadow while the net recovers), so drift cannot reach the
|
|
//--- account. No effect in the Strategy Tester/optimizer - the model is held fixed there by design.
|
|
//--- ON BY DEFAULT AND NO LONGER AN INPUT (2026-08-01). Adapting to a changing market is not an optional
|
|
//--- extra for a model that will be attached for months, it is the thing that keeps it from going stale,
|
|
//--- and the guardrail above is what makes it safe to leave on. See the caveat in the release checklist:
|
|
//--- this had never been forward-tested on a live feed at the time it was made default.
|
|
const bool EnableOnlineLearning = true;
|
|
//--- Default 6 -> 3 and NO LONGER AN INPUT (2026-08-01). Declustering existed because exact-pivot ZigZag
|
|
//--- labels make a same-direction repeat provably redundant; barrier labels answer every bar
|
|
//--- independently, so consecutive Buy setups inside a trend are real trades and suppressing them throws
|
|
//--- signal away - which argued for 0. It is not 0 because on D1 and above a 6-bar window spans more than
|
|
//--- a trading week, and two arrows a day apart on a weekly-scale move really are one event. 3 keeps the
|
|
//--- immediate-neighbour duplicate off the chart on slow timeframes while leaving genuine consecutive
|
|
//--- setups intact on fast ones. Display/emission only either way - the raw per-bar recall/precision
|
|
//--- metrics are never declustered, so this cannot flatter a model's measured numbers.
|
|
//--- 10 bars (2026-08-10, was 3). On H1 a 3-bar window collapsed only the tightest runs and left
|
|
//--- visible clusters around every turn; 10 bars is a third of a session and closer to the spacing of
|
|
//--- genuinely distinct setups on this timeframe. Applies to every topology - it is not per-network.
|
|
const int SignalClusterWindow = 10;
|
|
//--- EXCURSION-SIZE HEAD (Expert\AIBase\Excursion.mqh). A second small net that predicts how FAR price
|
|
//--- travels within the horizon - never which way, which is measured-closed on three instruments.
|
|
//--- STAGE 1 IS A MEASUREMENT: it trains beside the classifier and prints a Brier skill score against
|
|
//--- the constant base rate a fixed ATR multiple already assumes. It places no orders and moves no
|
|
//--- stops, so leaving it on costs only era time and leaving it off changes nothing else.
|
|
//--- NOT in the weights fingerprint: it is a separate network with its own weights, so it cannot alter
|
|
//--- the classifier's shape - the rule that keeps TRAIN_BATCH_SIZE out for the same reason.
|
|
const bool UseExcursionHead = true;
|
|
//--- Era cap. NOT AN INPUT as of 2026-08-01: it is a runaway backstop, not a training control. Training
|
|
//--- decides its own ending (the plateau ladder deploys the best checkpoint once escalation stops finding
|
|
//--- anything better), so in a healthy run this number is never reached and choosing it changes nothing;
|
|
//--- in an unhealthy one the useful response is to read the era log, not to raise a cap.
|
|
const MAX_ERAS_PRESET MaxErasPerRun = ME_1000;
|
|
//==================================================================================================
|
|
// AI INPUT FEATURES (the data the neural network sees each bar)
|
|
//==================================================================================================
|
|
input string AISignals = "AI Input Features"; // AI Input Features
|
|
//--- ind_Periods IS NO LONGER AN INPUT (2026-08-11). The number of bars per input sequence is now
|
|
//--- DERIVED - median confirmed swing leg over recent history, snapped to a coarse ladder and capped
|
|
//--- (see DeriveHistoryBars) - then pinned in the model's .cfg and ADOPTED on every later load, the
|
|
//--- same measure-once contract as the barrier geometry. Picking it by predictive skill instead was
|
|
//--- ruled out by the 2026-08-06 lag profile (no information at any lag 0-20): a best-of-N window
|
|
//--- scan would only ever mine noise. The ATR feature/barrier-unit lookback it also used to set is
|
|
//--- deliberately DECOUPLED and pinned at the old default below: the ATR indicator is created before
|
|
//--- the .cfg can be adopted, so deriving its period would let init ordering change the unit the
|
|
//--- pinned SL/TP multiples are expressed in.
|
|
#define ATR_FEATURE_PERIOD 20
|
|
input ENUM_APPLIED_VOLUME VolumeData = VOLUME_TICK; // Volume data type (tick / real)
|
|
input bool EnableVolume = true; // Feature: volume
|
|
input bool EnableTime = true; // Feature: time
|
|
input bool EnableATR = true; // Feature: volatility (ATR)
|
|
//--- MA/RSI as network input features, independent of the Classic Signals votes above (you can feed MA
|
|
//--- to the model without it voting, or vice versa). Uses PeriodMA/MA_Type/PeriodRSI (Classic Signals)
|
|
//--- as the starting period, then auto-tuned from there when Auto-tune indicators is on.
|
|
input bool EnableMAFeature = true; // Feature: Moving Average
|
|
input bool EnableRSIFeature = false; // Feature: RSI
|
|
//--- MACD adds 3 inputs/bar (main, signal, histogram - all ATR-normalized); Ichimoku adds 8 (distances to
|
|
//--- Tenkan/Kijun/both cloud edges, the TK spread, cloud thickness here and projected, and the Chikou
|
|
//--- displacement). Widths are per BAR, so each is multiplied by Bars to analyse before it reaches the
|
|
//--- first layer - Ichimoku at the default 20 bars is 160 extra inputs on its own. Worth it for the
|
|
//--- multi-timescale structure nothing else in the vector carries, but enable deliberately, not by habit.
|
|
input bool EnableMACDFeature = false; // Feature: MACD
|
|
input bool EnableIchimokuFeature = false; // Feature: Ichimoku
|
|
//--- Confirmed ZigZag swing direction/magnitude/age - lookahead-safe (repainting embargo applied, see
|
|
//--- SWING_CONFIRMATION_BARS in Expert\ExpertSignalAIBase.mqh).
|
|
input bool EnableSwingContext = true; // Feature: ZigZag swing context
|
|
input bool EnableADCumulativeDelta = false; // Feature: Cumulative Delta
|
|
input bool EnableADShorteningOfThrust = true; // Feature: Shortening of Thrust
|
|
input bool EnableADWyckoffEventStream = true; // Feature: Wyckoff Events
|
|
input bool EnableADWyckoffFailedStructure = true; // Feature: Wyckoff Failed Structure
|
|
input bool EnableADWyckoffSignificantBarInversion = true; // Feature: Wyckoff Bar Inversion
|
|
//==================================================================================================
|
|
// AD / WYCKOFF INDICATOR PARAMETERS
|
|
//==================================================================================================
|
|
//--- ADDED 2026-08-08, because AutoTuneIndicators now defaults to false and these 33 values had NO input
|
|
//--- of any kind - they were literals in CADIndicatorTuner's constructor. MA/RSI/MACD/Ichimoku have had
|
|
//--- their periods exposed since the beginning; the order-flow and Wyckoff indicators, which contribute
|
|
//--- 28 of the 64 features per bar on the AD configs, were operator-invisible. With the tuner on that was
|
|
//--- survivable (it searched them); with it off they would be frozen at values nobody chose.
|
|
//---
|
|
//--- SEEDS, exactly like PeriodMA/MA_Type. When AutoTuneIndicators is on, the search still starts here and
|
|
//--- is still free to move each indicator's copy independently - collapsing the shared thresholds below
|
|
//--- into one input each constrains only what the OPERATOR sets, never what the tuner may explore.
|
|
//---
|
|
//--- CONSOLIDATED 33 -> 18 on purpose, following the imbalance-input precedent. volClimax/volHigh/
|
|
//--- rangeClimax/rangeSignificant/stVolRatio/atr were duplicated verbatim across CumulativeDelta, Wyckoff
|
|
//--- Events, Failed Structure and Bar Inversion - the same four constants restated 3-4 times each. They
|
|
//--- are one CONCEPT per row ("what counts as climactic volume", "what counts as a significant range"),
|
|
//--- so they get one input per concept. Eighteen knobs an operator can reason about beats thirty-three
|
|
//--- that invite inconsistent settings for the same idea.
|
|
//---
|
|
//--- Every default below is byte-identical to the literal it replaces, so this ships as a pure no-op:
|
|
//--- see the ADP token in ConfigFingerprint(), which is appended ONLY when something actually differs,
|
|
//--- leaving every existing model's filename - and therefore its trained weights - untouched.
|
|
#define WYK_VOL_CLIMAX_DEF 2.5
|
|
#define WYK_VOL_HIGH_DEF 1.5
|
|
#define WYK_RANGE_CLIMAX_DEF 1.8
|
|
#define WYK_RANGE_SIGNIF_DEF 1.2
|
|
#define WYK_ST_VOL_RATIO_DEF 0.6
|
|
#define WYK_ATR_MULT_DEF 0.5
|
|
#define ADCD_LOOKBACK_DEF 50
|
|
#define SOT_THRUST_LOOKBACK_DEF 30
|
|
#define SOT_MIN_IMPULSES_DEF 3
|
|
#define SOT_THRESHOLD_DEF 0.30
|
|
#define WES_LOOKBACK_DEF 50
|
|
#define WES_ZIGZAG_DEF 3
|
|
#define WES_TOUCH_ATR_DEF 0.5
|
|
#define WES_AR_MIN_ATR_DEF 1.0
|
|
#define WES_MAX_RANGE_BARS_DEF 200
|
|
#define WFS_LOOKBACK_DEF 50
|
|
#define WFS_ZIGZAG_STRENGTH_DEF 3
|
|
#define WSBI_LOOKBACK_DEF 50
|
|
input string ADParams = "AD / Wyckoff Parameters"; // AD / Wyckoff Parameters
|
|
//--- Shared thresholds. Each is a multiple of that bar's own rolling average (volume) or ATR (range), so
|
|
//--- they are scale-free and mean the same thing on any symbol.
|
|
input double Wyk_VolClimaxMult = WYK_VOL_CLIMAX_DEF; // Climactic volume (x average)
|
|
input double Wyk_VolHighMult = WYK_VOL_HIGH_DEF; // High volume (x average)
|
|
input double Wyk_RangeClimaxMult = WYK_RANGE_CLIMAX_DEF; // Climactic range (x ATR)
|
|
input double Wyk_RangeSignificantMult = WYK_RANGE_SIGNIF_DEF; // Significant range (x ATR)
|
|
input double Wyk_ShortTermVolRatio = WYK_ST_VOL_RATIO_DEF; // Short-term volume ratio
|
|
input double Wyk_AtrMult = WYK_ATR_MULT_DEF; // General ATR multiple
|
|
//--- Per-indicator structure. Lookbacks are in BARS and set how far back each indicator searches for the
|
|
//--- structure it names; they do not need to agree with each other or with HistoryBars.
|
|
input int ADCD_Lookback = ADCD_LOOKBACK_DEF; // Cumulative Delta: lookback bars
|
|
input int SOT_ThrustLookback = SOT_THRUST_LOOKBACK_DEF; // Shortening of Thrust: lookback bars
|
|
input int SOT_MinImpulses = SOT_MIN_IMPULSES_DEF; // Shortening of Thrust: min impulses
|
|
input double SOT_Threshold = SOT_THRESHOLD_DEF; // Shortening of Thrust: threshold
|
|
input int WES_Lookback = WES_LOOKBACK_DEF; // Wyckoff Events: lookback bars
|
|
input int WES_ZigZag = WES_ZIGZAG_DEF; // Wyckoff Events: ZigZag strength
|
|
input double WES_TouchATR = WES_TOUCH_ATR_DEF; // Wyckoff Events: zone touch (x ATR)
|
|
input double WES_ARMinATR = WES_AR_MIN_ATR_DEF; // Wyckoff Events: min AR size (x ATR)
|
|
input int WES_MaxRangeBars = WES_MAX_RANGE_BARS_DEF; // Wyckoff Events: max range life (bars)
|
|
input int WFS_Lookback = WFS_LOOKBACK_DEF; // Failed Structure: lookback bars
|
|
input int WFS_ZigZagStrength = WFS_ZIGZAG_STRENGTH_DEF; // Failed Structure: ZigZag strength
|
|
input int WSBI_Lookback = WSBI_LOOKBACK_DEF; // Bar Inversion: lookback bars
|
|
//--- News LAST in this list on purpose: it is the only feature whose data comes from outside the price
|
|
//--- series (the terminal's economic calendar), so it is the one a user is most likely to want to reason
|
|
//--- about separately - and the only one with a companion setting. Proximity/impact only, never
|
|
//--- actual-vs-forecast, which is not knowable ahead of the release.
|
|
input bool EnableNews = false; // Feature: news proximity
|
|
input NF_LOOKBACK_PRESETS NewsFeatureWindowMinutes = M60; // News feature window
|
|
//--- Cross-asset, after News for the same reason: its data also comes from outside this symbol's own
|
|
//--- series - in fact it is the ONLY feature here that does so without leaving the price domain. Every
|
|
//--- other block above, News included, is either a transform of this one instrument's OHLCV or a
|
|
//--- timing overlay on it. Measured end to end, that whole family sits at the noise floor
|
|
//--- (research/test_classic.py), which is precisely why this exists. Builds a currency-strength panel
|
|
//--- from the FX pairs in Market Watch and feeds the traded pair's base/quote strength plus the
|
|
//--- divergence between the pair and its own two currencies. Needs >= 2 usable FX pairs in Market
|
|
//--- Watch; degrades to a neutral 0-fill with one logged line if it cannot build, never blocks training.
|
|
input bool EnableCrossAsset = true; // Feature: cross-asset currency strength
|
|
//--- Spread: the only microstructure channel that is both FX-available and genuinely historical in
|
|
//--- the Strategy Tester, so the only one a backtest can honestly validate. Encodes a volatility
|
|
//--- REGIME (spread is near-fixed while ATR is not, so the ratio runs high exactly when realised
|
|
//--- volatility is below its ATR estimate), which predicts whether ATR-scaled barriers get reached.
|
|
//--- Unsigned, like volume - it informs Neutral-vs-directional and can never pick a side.
|
|
input bool EnableSpreadFeature = true; // Feature: spread / volatility regime
|
|
//--- Searches the per-bar parameters of every ENABLED input feature above (the order-flow/Wyckoff
|
|
//--- MA/RSI feature periods) for the combination that trains best - see ADIndicatorTuner.mqh.
|
|
//--- The TRIAL COUNT IS NO LONGER AN INPUT (2026-08-01). It was a number the user had no basis to pick:
|
|
//--- the right budget depends on how many parameters are actually being searched and how wide each
|
|
//--- one's range is, both of which the code knows at runtime and the user does not. Asking for it
|
|
//--- guaranteed either a wasted search (too many trials on two narrow parameters) or a blind one (32
|
|
//--- trials against a space of millions). Now derived - see ComputeTuneTrialBudget().
|
|
//--- DEFAULT FLIPPED TO false 2026-08-08, and NOT because the search is broken - it is correct, and its
|
|
//--- own Sidak family-wise gate is what proves it. The search is simply not worth its cost on data at
|
|
//--- the MI noise floor, which is every instrument measured so far:
|
|
//--- * It scored 324 candidates per model on SP500 H1 and reported no improvement on ALL FOUR
|
|
//--- topologies (0.00236 -> 0.00236 on the three AD configs, 0.00370 -> 0.00370 on PAI), rejecting
|
|
//--- its own winner at selection p=1.0000. That is the gate working, not failing.
|
|
//--- * It cannot do better here by construction: it ranks candidates by MARGINAL mutual information,
|
|
//--- and the headline MI reads 0.00370 nats against a shuffled-label null of 0.00379 +/- 0.00061
|
|
//--- (p=0.4975). Every candidate is a noise draw, so the maximum over N of them is noise too, and
|
|
//--- the correction rejects it - more harshly the more candidates are tried.
|
|
//--- * The cost is not marginal: 2674s (CONV), 3271s (LSTM), 3372s (HYB) - 45 to 56 minutes per model,
|
|
//--- in ONE synchronous call with no yield, during which the chart is frozen and silent. It was also
|
|
//--- the amplifier for the indicator-handle leak (fixed, see ReInitADIndicators): 324 iterations x 6
|
|
//--- re-creates orphaned gigabytes and MT5 removed CONV and LSTM for running out of memory.
|
|
//--- Turn it ON deliberately, on an instrument whose per-feature MI actually clears its null, and expect
|
|
//--- to wait. The EA's own report says it plainest: "no per-feature indicator retuning will help".
|
|
//--- The input STAYS. The sweep is one function of twelve in AIBase\AutoTune.mqh - the other eleven are
|
|
//--- the MI/lag/excursion/geometry diagnostics that produced every verdict this project relies on, and
|
|
//--- they run on their own path regardless of this flag (see TuneIndicatorsAndTrain's else-branches).
|
|
input bool AutoTuneIndicators = false; // Auto-tune indicator params (slow)
|
|
//==================================================================================================
|
|
// FILTERS
|
|
//==================================================================================================
|
|
input string SF_Settings = "Session Filter"; // Session Filter
|
|
input bool EnableSessionFilter = false; // Signal: Session filter
|
|
//--- All three ON by default. The filter is evaluated once per BAR (Expert_EveryTick=false ships as the
|
|
//--- default), so on a slow timeframe there are very few evaluations per day and a single-session
|
|
//--- default can starve the EA of entries entirely - on D1 there is exactly ONE evaluation, at the bar
|
|
//--- open, and whether that instant falls inside a narrow session window depends purely on the broker's
|
|
//--- server offset. Enabling all three spans 00:00-22:00 GMT so only genuinely dead hours are excluded;
|
|
//--- narrow it deliberately per-chart rather than inheriting it as an accident of the default.
|
|
input bool SF_trade_LondonSession = true; // Trade London session
|
|
input bool SF_trade_TokyoSession = true; // Trade Tokyo session
|
|
input bool SF_trade_NewYorkSession = true; // Trade New York session
|
|
//--- Scheduled flat-close. Deliberately its OWN group rather than part of the Session Filter above:
|
|
//--- CExpertCustom::OnTick() (Expert\ExpertCustom.mqh) evaluates this schedule unconditionally, so it
|
|
//--- fires whether EnableSessionFilter is on or off - grouping it under the session filter implied a
|
|
//--- coupling that has never existed in the code. Set Close-all day = Disabled to switch it off.
|
|
input string CA_Settings = "Scheduled Close-All"; // Scheduled Close-All
|
|
input CLOSE_DAY_OF_WEEK targetDayOfWeek = CLOSE_FRIDAY; // Close-all day
|
|
input CLOSE_HOUR_OF_DAY targetHour = CH_23; // Close-all hour
|
|
input CLOSE_MINUTE_OF_HOUR targetMinutes = CM_45; // Close-all minute
|
|
//--- INTRADAY TIME FILTER REMOVED ENTIRELY 2026-08-01 (5 inputs, plus Signals\SignalITF.mqh).
|
|
//--- Two of its five inputs were raw BITMASKS ("hours to avoid" as an integer), which is not a setting a
|
|
//--- trader can reasonably compute - it is an implementation detail exposed as a control, and it shipped
|
|
//--- disabled so essentially nobody ever got it right. More importantly the job is now covered three
|
|
//--- times over by things that learn or are declarative: the Session Filter handles "when may I trade"
|
|
//--- explicitly, the time-of-day/day-of-week features (EnableTime) let the NETWORK discover which hours
|
|
//--- are good on this instrument instead of being told, and the trade journal ranks by time bucket.
|
|
//--- A hand-specified hour mask is the least informed of the four and the hardest to use.
|
|
input string NF_Settings = "News Filter"; // News Filter
|
|
input bool EnableNewsFilter = true; // Signal: News filter
|
|
input NF_LOOKBACK_PRESETS NF_LookMinutes = M60; // News avoid window (min)
|
|
input NF_IMPACT_PRESETS NF_MinImpact = HOLIDAYS; // Min news impact to avoid
|
|
//--- MARKET DEPTH FILTER REMOVED ENTIRELY 2026-08-01 (5 inputs, plus Signals\SignalMarketDepth.mqh).
|
|
//--- It needs real level-2 DOM data, which this development broker does not provide and which most
|
|
//--- retail MT5 brokers do not provide either - so the module has never been executed against real data
|
|
//--- even once. Shipping four tuning dropdowns for an UNTESTED code path is worse than shipping nothing:
|
|
//--- the only users who could enable it are the ones whose broker supplies a book, and they would be
|
|
//--- the first people ever to run it, in live trading, with no validation behind it. If DOM support is
|
|
//--- wanted later it should return as a feature fed to the network rather than as a rule-based veto with
|
|
//--- its own hand-tuned thresholds - the imbalance is data, and data belongs in the input vector.
|
|
input string RiskGuard_Settings = "Risk Guard"; // Risk Guard
|
|
input bool EnableRiskGuard = true; // Signal: Risk Guard
|
|
//--- FREE-ENTRY PERCENTAGES, replacing the RISK_LIMIT_PCT_PRESET dropdown these two used to be
|
|
//--- (that enum is gone - see Enumerations\InputEnums.mqh). Every funded/prop programme sets its own
|
|
//--- numbers and they are not always integers, so a fixed ladder of presets could not express them;
|
|
//--- 4.5% or 3.75% were simply unreachable. Enter the limits from YOUR account agreement, and enter
|
|
//--- them slightly TIGHTER than the contract if you want margin for slippage past a stop.
|
|
//--- 0 disables a rule. Enforced live at quote frequency by Variables\RiskBudget.mqh - not once per
|
|
//--- bar, which is all the old guard could manage.
|
|
input double MaxDailyLossPct = 4.0; // Daily loss limit % (0 = off)
|
|
input double MaxDrawdownPct = 8.0; // Max total drawdown % (0 = off)
|
|
//--- TRUE: max drawdown is measured down from the highest equity ever reached (trailing DD, the
|
|
//--- stricter and more common funded-account rule). FALSE: measured from the equity this EA first
|
|
//--- saw on the account (static DD). Pick whichever your programme actually uses - a trailing rule
|
|
//--- applied to a static challenge halts trading long before it has to.
|
|
input bool MaxDrawdownIsTrailing = true; // Max DD trails the equity peak
|
|
//--- Broker-server hour at which the firm's trading day (and therefore the daily loss allowance)
|
|
//--- resets. Broker time here is NOT your local time; if the firm quotes the reset in another zone,
|
|
//--- convert it. A misaligned window hands the allowance back hours early or late.
|
|
input int RiskDayResetHour = 0; // Risk day reset hour (broker time, 0-23)
|
|
//--- Ceiling on what ONE trade may risk, as a share of the allowance that is genuinely left after
|
|
//--- subtracting every open position's remaining loss-to-stop. This is the fix for the real breach
|
|
//--- mode: without it a trade at 3.2% into a 4% day still sized for a full risk unit and a routine
|
|
//--- stop-out went through the limit. At the default 50% a full stop-out spends at most half of
|
|
//--- what is left, so even a stop that slips to twice its distance lands inside the limit.
|
|
input double RiskPerTradeOfBudget = 50.0; // Max % of remaining budget per trade
|
|
//--- Blocking new entries cannot stop an ALREADY-OPEN position from running through the limit, which
|
|
//--- is the way a hard daily loss rule is actually breached. Turn this on to close this EA's own
|
|
//--- positions (matching symbol + magic) the moment a limit is hit. Default OFF because closing
|
|
//--- positions is a materially bigger behaviour change than declining to open them - but leaving it
|
|
//--- off means the limits above are advisory, not enforced.
|
|
input bool RiskGuardFlatten = false; // Close own positions on breach
|
|
//--- EXPECTANCY STOP. The daily and total limits bound how FAST the account can lose; neither notices
|
|
//--- WHETHER it is losing. A negative-expectancy signal traded inside a 4%/8% envelope breaches no rule
|
|
//--- and still arrives at zero - it just takes longer. This tests the realised mean result per trade
|
|
//--- against zero and stops opening new positions once it is significantly below.
|
|
//--- Significantly, not merely below: a run of losers is ordinary variance even for a profitable system,
|
|
//--- so the test uses the standard error of the mean and a wide spread simply demands more trades before
|
|
//--- it can fire. Measured in R (net profit over money risked), so symbols and lot sizes share one scale,
|
|
//--- and net of swap and commission - when the directional edge is zero, cost IS the expectancy.
|
|
//--- 0 trades = off. The halt is LATCHED and survives a restart; clearing it means deleting the risk
|
|
//--- state file, deliberately, after looking at why.
|
|
input int ExpectancyMinTrades = 40; // Halt if losing: min closed trades first (0 = off)
|
|
input double ExpectancySigma = 2.0; // ...and mean must be this many std errors below zero
|
|
//==================================================================================================
|
|
// TRADE JOURNAL / PATTERN RANKING
|
|
//==================================================================================================
|
|
input string Journal_Settings = "Trade Journal / Ranking"; // Trade Journal / Ranking
|
|
//--- Enables the per-pattern win-rate database: scales each signal's vote by its historical win rate,
|
|
//--- records every trade, and powers the Export Trade Journal Report button (see the control panel).
|
|
//--- Default TRUE since 2026-08-13 (user request): a META chart should journal + rank out of the box,
|
|
//--- and the classic-only configuration benefits from ranked weights as soon as history accumulates.
|
|
input bool UseDatabaseRanking = true; // Weight filters by DB win-rate
|
|
//--- Row cap per pattern table (oldest row pruned past it). 1000 is plenty for live ranking; a
|
|
//--- META-LABEL CORPUS BUILD (Meta_Labeling_Design.md, stage S1: a long backtest whose signal DB
|
|
//--- becomes the training set) needs it raised so a 15-20 year run isn't pruned away - 20000 holds
|
|
//--- ~3x the densest pattern's 20-year stretch count. A high cap costs nothing until rows exist.
|
|
input int DB_MaxRowsPerTable = 1000; // Max rows kept per pattern table
|
|
//--- META DATASET EXPORT (cross-sectional pooling, Meta_Labeling_Design.md). With AIType=META, the
|
|
//--- chart writes its full training set - every resolved candidate's feature window + setup
|
|
//--- descriptor + triple-barrier label - to Common\Files\Warrior_EA\MetaExport\<sym>_<period>.f32
|
|
//--- (float32 rows; sidecar .meta.csv carries width/geometry/BE) once per attach, then trains as
|
|
//--- normal. Pooled training across symbols happens OFFLINE on these files; the EA itself is
|
|
//--- unchanged. Costs one pass-1-sized sweep (~a minute) at attach. Default ON in the private build
|
|
//--- (the pooling campaign's drop-on-chart workflow); OFF for Market.
|
|
#ifdef WARRIOR_MARKET_BUILD
|
|
input bool Meta_ExportDataset = false; // META: export training dataset at attach
|
|
#else
|
|
input bool Meta_ExportDataset = true; // META: export training dataset at attach
|
|
#endif
|
|
//--- Header only. The AI\Network.mqh optimizer inputs (Adam*, Sgd*) are declared in that library
|
|
//--- header; because this Inputs file is included FIRST (see Warrior_EA.mq5), those
|
|
//--- render immediately AFTER this divider - grouping them here instead of leading the Inputs tab.
|
|
input string NNPerf_Settings = "NN Optimizer / Performance"; // NN Optimizer / Performance
|