forked from animatedread/Warrior_EA
a303f5b86c
Eliminate the separate `AI_TOPOLOGY_PRESET` enum and input. Fold the topology presets directly into `AI_CHOICE` as new combined values (MLP_3L, MLP_4L, CONV_2L, LSTM_2L, HYBRID_2L) plus `AI_NONE`. Remove the `TopologyPreset` input variable and update default `AIType` assignments. Update the market description to reflect the simplified single‑selector interface. **Why:** Users previously had to choose an AI architecture and a topology preset separately. Now the UI shows one coherent selector that bundles architecture with its appropriate dense‑layer depth, reducing complexity and preventing mismatches.
668 lines
No EOL
28 KiB
HTML
668 lines
No EOL
28 KiB
HTML
<h2>Introducing Warrior EA – A Self‑Training Neural Network for Any Market</h2>
|
|
|
|
<p><strong>Warrior EA</strong> trains its own neural network directly on the chart you attach it to. It learns from your instrument’s history, evaluates itself on unseen data, and only trades once the model demonstrates reliable behavior. The EA supports <u>any symbol and any timeframe</u>, adapting naturally to different market conditions.</p>
|
|
|
|
|
|
<h2>Trading Strategy</h2>
|
|
|
|
<p>The EA can operate using its self‑training AI or a classic multi‑indicator path (MA, RSI, MACD, Ichimoku). The neural network reads price action, volatility, volume, cyclical time structure and confirmed swing structure, forming Buy, Sell or Neutral classifications.</p>
|
|
<p>A portion of history is held back during training, ensuring that accuracy is measured only on data the model has not seen. When both AI and classic paths are enabled, they vote together for stable decision‑making.</p>
|
|
|
|
|
|
<h2>Key Features</h2>
|
|
|
|
|
|
<ol>
|
|
<li>Self‑Training Neural Network: Perceptron, Convolutional, LSTM or Hybrid architectures.</li>
|
|
<li>Unseen‑Data Validation: The model must reliably recognize Buy, Sell and Neutral before going live.</li>
|
|
<li>Adaptive Learning: The deployed model continues adjusting to newly confirmed structure.</li>
|
|
<li>Wyckoff Context: Optional structural readings for accumulation, distribution, exhaustion and control shifts.</li>
|
|
<li>Classic Signals: MA, RSI, MACD and Ichimoku paths available while the AI warms up.</li>
|
|
<li>No External Software: Native MQL5/OpenCL only — runs unchanged on any VPS or cloud agent.</li>
|
|
</ol>
|
|
|
|
|
|
<h2>Risk Management</h2>
|
|
|
|
<p>Stops anchor to the recent swing extreme, with optional volatility padding or adaptive distance. Position size is calculated from the actual stop distance so the chosen risk percentage is always respected.</p>
|
|
<p>Targets can use volatility multiples, the opposite swing or an adaptive reward based on confidence. Daily‑loss and drawdown limits are available for disciplined trading.</p>
|
|
|
|
|
|
<h2>Trade Journal & Self‑Ranking Database</h2>
|
|
|
|
<p>Every closed trade is written to a local database: entry and exit price, stop and target, lot size, real profit/loss, maximum adverse/favorable excursion (in R‑multiples), AI confidence, database confidence, exit reason and the triggering pattern.</p>
|
|
<p>This runs automatically once “Weight filters by DB win‑rate” is enabled.</p>
|
|
<p>The same history powers a self‑ranking system for classic signals: each pattern’s win rate is tracked over time and its influence is scaled accordingly, so patterns that actually work for your symbol and timeframe gain more weight.</p>
|
|
<p>The Report button exports the full history to CSV: overall win rate, average R‑multiple, breakdown by hour/day, confidence bracket, and plain‑language suggestions based on recurring weak spots. Every raw trade is included for deeper analysis.</p>
|
|
|
|
|
|
<h2>Benefits</h2>
|
|
|
|
|
|
<ul>
|
|
<li>Works on Forex, indices, metals, commodities and crypto.</li>
|
|
<li>No curve‑fitting — features are scale‑normalized across instruments.</li>
|
|
<li>Clear on‑chart feedback: training progress, unseen‑data accuracy and live signal reliability.</li>
|
|
<li>Self‑improving classic signals weighted by their own tracked win rate.</li>
|
|
</ul>
|
|
|
|
|
|
<h2>Internal Workflow (Summary)</h2>
|
|
|
|
|
|
<ul>
|
|
<li>Collects features over the chosen lookback window.</li>
|
|
<li>Trains the selected neural architecture using Adam or SGD.</li>
|
|
<li>Validates on the OOS split and selects the best-performing era.</li>
|
|
<li>Deploys the model and begins adaptive learning (if enabled).</li>
|
|
</ul>
|
|
|
|
|
|
<h2>Notes</h2>
|
|
|
|
|
|
<ul>
|
|
<li>Performance depends on broker data quality and symbol behavior.</li>
|
|
<li>Training time varies by hardware and chart history length.</li>
|
|
<li>Online learning is disabled in Strategy Tester by design.</li>
|
|
</ul>
|
|
|
|
|
|
<h2>FAQ</h2>
|
|
|
|
<p><strong>Does the EA need optimization?</strong> AI: No — defaults are tested. Classic signals (MA, RSI, MACD, Ichimoku) may benefit from optimization depending on symbol and timeframe.</p>
|
|
<p><strong>Does it repaint?</strong> No — all structure is confirmed before training.</p>
|
|
<p><strong>Can it trade immediately?</strong> Yes — via classic signals (MA, RSI, MACD, Ichimoku) while the AI warms up.</p>
|
|
|
|
|
|
<h2>Results</h2>
|
|
|
|
<p>The EA trains a separate model per symbol and timeframe, so performance depends on your own chart and broker. Evaluate it in the Strategy Tester, review unseen‑data accuracy, and export the journal for detailed analysis.</p>
|
|
<p>Trading involves risk. Test on a demo account before live use.</p>
|
|
|
|
|
|
<h2>Get Started</h2>
|
|
|
|
<p>Attach the EA to the symbol and timeframe you intend to trade, use a different Magic Number per chart, and enable the AI to allow the warm‑up period. The defaults are ready to trade. To enable the exportable trade journal, turn on “Weight filters by DB win‑rate”.</p>
|
|
|
|
|
|
<h2>Input Parameters</h2>
|
|
|
|
<p>Every input includes an on‑screen description. The defaults are the tested configuration.</p>
|
|
|
|
|
|
<table cellspacing="0" cellpadding="2" width="100%" class="standart">
|
|
<thead>
|
|
<tr>
|
|
<th>Variable Name</th>
|
|
<th>Parameter Name</th>
|
|
<th>Parameter Description</th>
|
|
<th>Default</th>
|
|
</tr>
|
|
</thead>
|
|
|
|
|
|
<tbody>
|
|
<tr>
|
|
<td colspan="4"><strong>General</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Expert_MagicNumber</td>
|
|
<td>Magic number</td>
|
|
<td>Identifies this EA's own trades so it never confuses them with another EA or a manual position. Use a different value per chart.</td>
|
|
<td>2024</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Expert_EveryTick</td>
|
|
<td>Calculate on every tick</td>
|
|
<td>Recalculates signals every tick, or only once a new bar forms. Affects both results and CPU load.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>VerboseMode</td>
|
|
<td>Detailed panel + verbose journal</td>
|
|
<td>Off: plain-language panel and a quiet log. On: full technical detail and extra diagnostics, for studying the model.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Money Management</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>MM_STRATEGY / Money_FixLot_Lots</td>
|
|
<td>MM strategy / fixed lot</td>
|
|
<td>Fixed risk %, Intelligent or Fixed lot. Both risk-based modes size from the trade's real stop distance, so the Risk % below is what is actually risked; Intelligent scales it down (never up) by a quarter-Kelly estimate. Fixed lot trades the size set here.</td>
|
|
<td>Fixed risk % / 0.01</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Money_Risk_Percent</td>
|
|
<td>Risk % of balance per trade</td>
|
|
<td>Percentage of balance risked per trade. Lot size derives from this and the actual stop distance, so the figure holds whichever stop mode is chosen. Also the ceiling Intelligent sizing scales down from.</td>
|
|
<td>1%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Trade Management</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>tradingdirection</td>
|
|
<td>Trade direction</td>
|
|
<td>Restricts the EA to buys only, sells only, or both.</td>
|
|
<td>Both</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Entry_Multiplier</td>
|
|
<td>Entry type/offset</td>
|
|
<td>Where the order sits relative to price: at market, a limit or stop offset by 1-3x volatility, at the last swing, or an AI-confidence-scaled pullback.</td>
|
|
<td>Market</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>SL_Mode</td>
|
|
<td>Stop-loss mode</td>
|
|
<td>How far beyond the recent swing the stop sits. The anchor is always the swing extreme over the Bars to analyse window - lowest low for a buy, highest high for a sell - and this chooses only the padding: a fixed volatility multiple, none, or an adaptive pad tightening as conviction rises. A minimum distance from entry is always enforced.</td>
|
|
<td>1 x ATR from previous swing</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>TP_Mode</td>
|
|
<td>Take-profit mode</td>
|
|
<td>Where the target sits. ATR modes measure a volatility multiple from entry; Previous swing targets the opposite swing; Intelligent targets a multiple of that trade's own risk - 2.5x at zero confidence to 5x at full conviction - keeping reward:risk coherent with the stop.</td>
|
|
<td>Previous opposed swing</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Min_Risk_Reward_Ratio</td>
|
|
<td>Min reward:risk (reject only)</td>
|
|
<td>A trade is skipped if its reward:risk falls below this. It only rejects; it never resizes stop or target.</td>
|
|
<td>1:2</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>TrailingStrategy</td>
|
|
<td>Trailing stop</td>
|
|
<td>How the stop trails an open position once price moves in its favour.</td>
|
|
<td>None</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Signal_Expiration</td>
|
|
<td>Pending order expiry (bars)</td>
|
|
<td>How many bars a pending order stays active before being cancelled.</td>
|
|
<td>3 bars</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Confidence_Source</td>
|
|
<td>AI confidence source</td>
|
|
<td>Which confidence feeds the adaptive stop, target, trailing, early exit and Intelligent sizing.</td>
|
|
<td>AI only</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Min_Vote_Open</td>
|
|
<td>Min vote to open (AI + classic)</td>
|
|
<td>One unified gate for both engines, on a 0-100 scale: how much combined signal agreement is required to open a trade. Classic filters contribute their pattern weight; the AI contributes the weight of its confidence tier - 25, 50, 75 or 100 - so a hesitant call votes weakly and a near-certain one votes at full strength. There is no separate AI confidence setting: raising this demands a more confident AI in exactly the way it demands stronger classic agreement.</td>
|
|
<td>20%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Min_Vote_Close</td>
|
|
<td>Min opposite vote to close (AI + classic)</td>
|
|
<td>One unified gate for both engines: how much combined signal agreement in the opposite direction is required to close an open trade, and how strongly the AI must have turned against the position before that alone closes it. Kept above Min vote to open so leaving a trade demands more conviction than entering it, and a brief flicker against the position does not close it. Set it to Disabled to switch off signal-driven exits entirely, leaving stop-loss, take-profit and trailing.</td>
|
|
<td>80%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Classic Signals</strong> - the no-training fallback path</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableMA / EnableRSI</td>
|
|
<td>MA / RSI classic vote</td>
|
|
<td>Turn the Moving Average and RSI signals on or off.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>PeriodMA / PeriodRSI</td>
|
|
<td>MA / RSI period</td>
|
|
<td>Periods used by the two classic indicators. Shared by the vote and the matching AI input feature.</td>
|
|
<td>50 / 14</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>MA_Type</td>
|
|
<td>MA type</td>
|
|
<td>Averaging method - SMA, EMA, SMMA, LWMA, ALMA, DEMA, ZLEMA, T3 or Kalman. Shared by the vote and the MA input feature.</td>
|
|
<td>EMA</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableMACD / EnableIchimoku</td>
|
|
<td>MACD / Ichimoku classic vote</td>
|
|
<td>Turn the MACD and Ichimoku signals on or off. MACD adds momentum and price/oscillator divergence; Ichimoku adds multi-timeframe support and resistance structure - cloud position, line crossings, breakouts and lagging-span confirmation.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>MACD_PeriodFast / MACD_PeriodSlow / MACD_PeriodSignal</td>
|
|
<td>MACD periods</td>
|
|
<td>Periods for the fast EMA, slow EMA, and signal line in MACD. Controls the momentum and divergence sensitivity: faster periods react quicker to price changes, slower periods smooth out noise. Every combination offered is valid, and they are shared by the vote and the matching AI input feature.</td>
|
|
<td>12 / 26 / 9</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>Ichimoku_PeriodTenkan / Ichimoku_PeriodKijun / Ichimoku_PeriodSenkou</td>
|
|
<td>Ichimoku periods</td>
|
|
<td>Periods for Tenkan-sen (conversion line, short-term midpoint), Kijun-sen (base line, medium-term midpoint), and Senkou Span B (long-term midpoint forming the cloud). Controls multi-timeframe support and resistance: shorter Tenkan/Kijun periods react quicker, longer Senkou B creates a wider cloud. Every combination offered is valid, and they are shared by the vote and the matching AI input feature.</td>
|
|
<td>9 / 26 / 52</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>AI Input Features</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>ind_Periods</td>
|
|
<td>Bars to analyse</td>
|
|
<td>Lookback window for the AI's features, and the swing window the stop anchors to.</td>
|
|
<td>20 bars</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>VolumeData</td>
|
|
<td>Volume data type</td>
|
|
<td>Which volume series is used wherever volume is read.</td>
|
|
<td>Tick volume</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableVolume</td>
|
|
<td>Feature: volume</td>
|
|
<td>Includes volume as change against the previous bar, so it means the same on any instrument.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableTime</td>
|
|
<td>Feature: time</td>
|
|
<td>Adds 6 cyclical readings: time of day, day of week and month of year, each as a sine/cosine pair so 23:00 and 00:00, or December and January, are adjacent rather than opposite extremes. Lets the model read session rhythm, weekday effects and seasonality.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableATR</td>
|
|
<td>Feature: volatility (ATR)</td>
|
|
<td>Includes volatility as a fraction of price, so it stays comparable across instruments and years.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableMAFeature / EnableRSIFeature</td>
|
|
<td>Feature: Moving Average / RSI</td>
|
|
<td>Feeds the MA and RSI to the network as inputs, independently of whether they also vote as classic signals.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableMACDFeature / EnableIchimokuFeature</td>
|
|
<td>Feature: MACD / Ichimoku</td>
|
|
<td>Feeds MACD and Ichimoku to the network as inputs, independently of whether they also vote as classic signals. MACD adds 3 readings, Ichimoku 8.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableSwingContext</td>
|
|
<td>Feature: ZigZag swing context</td>
|
|
<td>Confirmed swing structure - direction, size, age and retracement of the last swing - plus where price sits in its recent range and how far it has stretched from its average.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableNews / NewsFeatureWindowMinutes</td>
|
|
<td>Feature: news proximity</td>
|
|
<td>Proximity to a news event - not the outcome, which isn't knowable ahead of time - over a configurable window.</td>
|
|
<td>False / 60 minutes</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableADCumulativeDelta</td>
|
|
<td>Feature: Cumulative Delta</td>
|
|
<td>Adds 6 order-flow readings: pressure, cumulative delta, the buy- and sell-side halves separately, absorption (effort without result) and initiative (effort with result).</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableADShorteningOfThrust</td>
|
|
<td>Feature: Shortening of Thrust</td>
|
|
<td>Adds 4 exhaustion readings: how much shorter the latest same-direction thrust is than the one before, whether it cost more volume, how many pushes have stacked up, and one firing only once a counter-move confirms it.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableADWyckoffEventStream</td>
|
|
<td>Feature: Wyckoff Events</td>
|
|
<td>Adds 12 readings: the current named event (all 14 - PS, SC, AR, ST, Spring, LPS, SOS plus PSY, BC, ARB, STB, UTAD, LPSY, SOW), the range's boundaries as volatility-normalised distance from price, the structural phase reached, change-of-character both ways, four sloped-structure classes and reaccumulation/redistribution flags.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableADWyckoffFailedStructure</td>
|
|
<td>Feature: Wyckoff Failed Structure</td>
|
|
<td>Adds 5 readings across two failure families: a swing turning before it reaches the structural high/low, and a boundary break rejected on the retest - each with a signed quality score.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableADWyckoffSignificantBarInversion</td>
|
|
<td>Feature: Wyckoff Bar Inversion</td>
|
|
<td>Adds 5 readings for climactic control-transfer bars: wide range and high volume against recent averages, closing at an extreme, clearing a volatility gate and breaking the last opposing significant bar - plus the control-flip case.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>AutoTuneIndicators</td>
|
|
<td>Auto-tune indicator params</td>
|
|
<td>Searches the best parameter values for whichever features are enabled - lookback windows, volume and range multiples, swing sensitivity, periods. Runs during training only; they lock on deployment so the win-rate history stays valid.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>IndicatorTuneTrials</td>
|
|
<td>Auto-tune trials</td>
|
|
<td>How many combinations the search evaluates. Weak candidates are discarded early, so a higher number costs less than it looks.</td>
|
|
<td>8</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Session Filter</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableSessionFilter / SF_trade_LondonSession / SF_trade_TokyoSession / SF_trade_NewYorkSession</td>
|
|
<td>Session filter</td>
|
|
<td>Restricts trading to the chosen sessions. All three ship on, covering 00:00-22:00 GMT - the filter is evaluated once per bar, so a single-session setting can starve a slow timeframe of entries. Does not affect Scheduled Close-All.</td>
|
|
<td>On, all three</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Scheduled Close-All</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>targetDayOfWeek / targetHour / targetMinutes</td>
|
|
<td>Close-all day/hour/minute</td>
|
|
<td>Closes all positions and pending orders at the chosen day, hour and minute - typically the weekend close. Runs independently of the session filter. Day = Disabled switches it off, Every Day makes it daily.</td>
|
|
<td>Friday 23:45</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Intraday Time Filter</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableITF / ITF_GoodHourOfDay / ITF_BadHoursOfDay / ITF_GoodDayOfWeek / ITF_BadDaysOfWeek</td>
|
|
<td>Intraday time filter</td>
|
|
<td>Marks a preferred hour and day, and hours or days to avoid, for finer time-of-day filtering.</td>
|
|
<td>On, none set</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>News Filter</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableNewsFilter</td>
|
|
<td>Signal: News filter</td>
|
|
<td>Turns the news filter on or off.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>NF_LookMinutes / NF_MinImpact</td>
|
|
<td>News window and impact</td>
|
|
<td>Minutes before and after an event during which no trade opens, and the minimum impact level that triggers it.</td>
|
|
<td>60 minutes / Holidays</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Market Depth Filter</strong> - live trading only</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableMarketDepth</td>
|
|
<td>DOM imbalance filter (live)</td>
|
|
<td>Order-book imbalance check that can confirm or veto a signal. Needs a broker providing real depth-of-market data.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>DOM_DepthLevels / DOM_ImbalanceScale / DOM_MaxSpreadMultiple</td>
|
|
<td>DOM levels, influence, spread cap</td>
|
|
<td>Order-book levels per side, how strongly the imbalance may influence the signal, and the spread multiple above which a trade is vetoed.</td>
|
|
<td>5 / 100% / 3x</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Risk Guard</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableRiskGuard</td>
|
|
<td>Signal: Risk Guard</td>
|
|
<td>Turns the daily-loss/drawdown halt on or off.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>MaxDailyLossPct / MaxDrawdownPct</td>
|
|
<td>Max daily loss / drawdown %</td>
|
|
<td>Once the day's loss, or drawdown from peak, reaches these percentages, no new trades open until the limit resets. Open positions are never force-closed.</td>
|
|
<td>Disabled</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Trade Journal / Ranking</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>UseDatabaseRanking</td>
|
|
<td>Weight filters by DB win-rate</td>
|
|
<td>Tracks how each signal pattern has performed in this EA's own history and shifts influence toward what has worked. Also enables the journal the Report button exports - turn this on to use the trade journal.</td>
|
|
<td>False</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>Neural Network</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>AIType</td>
|
|
<td>AI algorithm</td>
|
|
<td>Selects which network trains and trades - Perceptron, Convolutional, LSTM or Hybrid - or Disabled, falling back to MA/RSI alone. The default is Disabled in the Market build and Hybrid in the private build so a fresh install either trades immediately or starts with the full AI stack.</td>
|
|
<td>Disabled / Hybrid</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>TrainingOptimizer / OutputNeuronsCount</td>
|
|
<td>Optimizer and output type</td>
|
|
<td>Which algorithm updates the weights, and whether it trains as a 3-way Buy/Sell/Neutral classifier or a single-value regression.</td>
|
|
<td>Adam / Classification</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>InitialNeurons / MinNeuronsCount / NeuronsReduction / LstmHiddenSize / ConvFilterCount / ConvPoolWindow / ConvPoolStep</td>
|
|
<td>Network shape</td>
|
|
<td>Network topology: neurons in the first layer, the floor the tapering layers may not shrink below, the percentage each layer shrinks by, and the memory, filter, pooling-window and pooling-step sizes used by the LSTM, Convolutional and Hybrid types. The AI selector itself now chooses the preset: MLP 3L / 4L, CONV 2L, LSTM 2L, Hybrid 2L or Disabled.</td>
|
|
<td>500 / 20 / 70% / 32 / 16 / 3 / 2</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>StudyPeriods</td>
|
|
<td>Training years</td>
|
|
<td>How many years of chart history the network trains on.</td>
|
|
<td>10 years</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>OOSSplit</td>
|
|
<td>Out-of-sample holdout</td>
|
|
<td>Share of history held back and never trained on, used only to measure unseen-data performance.</td>
|
|
<td>30%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>MinRecall</td>
|
|
<td>Min per-class OOS recall %</td>
|
|
<td>How well each of Buy, Sell and Neutral must be recognised on held-back data before a model may go live, so it cannot ship by ignoring one direction. This is 3-way classification, not trade win rate - random guessing scores about 33%. There is no accuracy target: training runs until it stops improving, then deploys its best model.</td>
|
|
<td>40%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>AILogitPriorStrength</td>
|
|
<td>AI: prior correction to true base rate</td>
|
|
<td>Turning points are rare, so the network trains on balanced data and would otherwise call too many Buys and Sells. This re-weights each answer by how often it truly occurs, so live trading performs more like the training figures. Higher = stronger correction toward the true base rate.</td>
|
|
<td>25%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>OversampleParity</td>
|
|
<td>AI: oversample parity</td>
|
|
<td>The precision/frequency dial: lower gives fewer but more selective Buy/Sell calls, higher gives more. Neutral stays deliberately heavier, so this mostly governs how aggressively the model calls the directional classes.</td>
|
|
<td>60%</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>FocalLossGamma</td>
|
|
<td>Focal-loss gamma</td>
|
|
<td>How strongly training concentrates on hard-to-classify bars. 0 disables it in favour of plain class balancing, while 1.0 is the lighter default used here.</td>
|
|
<td>1.0</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>SwingConfirmationBars</td>
|
|
<td>Swing confirm bars (label)</td>
|
|
<td>Bars to wait after a candidate swing reversal before trusting it for training, since recent swings can still change shape.</td>
|
|
<td>100 bars</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>EnableOnlineLearning</td>
|
|
<td>AI: keep learning live from confirmed bars</td>
|
|
<td>After deployment, keeps adapting the model on a live chart to newly confirmed structure, only while recent accuracy holds up. No effect in the Strategy Tester.</td>
|
|
<td>True</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>SignalClusterWindow / MaxErasPerRun</td>
|
|
<td>Decluster window / max eras</td>
|
|
<td>Collapses consecutive same-direction arrows within this many bars to one (display only), and caps how many training passes run before the EA checks in.</td>
|
|
<td>6 / 1000</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td colspan="4"><strong>NN Optimizer / Performance</strong></td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>AdamLearningRate / AdamBeta1 / AdamBeta2</td>
|
|
<td>Adam optimizer settings</td>
|
|
<td>Learning rate and decay rates for the Adam optimizer (used when Optimizer = Adam). The defaults are the standard published values and rarely need changing.</td>
|
|
<td>0.0003 / 0.9 / 0.999</td>
|
|
</tr>
|
|
|
|
|
|
<tr>
|
|
<td>SgdLearningRate / SgdMomentum</td>
|
|
<td>SGD optimizer settings</td>
|
|
<td>Learning rate and momentum for the SGD+momentum optimizer (used when Optimizer = SGD).</td>
|
|
<td>0.0003 / 0.9</td>
|
|
</tr>
|
|
</tbody>
|
|
</table>
|
|
|
|
|
|
<h2>Final Notes</h2>
|
|
|
|
<p>Warrior EA is designed for traders who want adaptive automation without external dependencies or manual optimization. Review the training panel, check unseen‑data accuracy, and explore the trade journal for deeper insights.</p> |