/// <summary>Returns the feature string for the candle at position 'neg_idx'</summary>
public string FeaturesString(NegIdx neg_idx)
{
m_sb = m_sb ?? new StringBuilder();
m_sb.Clear();
// Output feature values
m_sb.Append("|features ");
foreach (var v in Features)
{
m_sb.Append(" ").Append(v.Value);
}
return(m_sb.ToString());
}
/// <summary></summary>
public override void Step()
{
// Avoid boundary cases in the data
if (Instrument.Count < PredictorNeuralNet.HistoryWindow + WindowSize)
{
return;
}
// Only do this per candle
if (!Instrument.NewCandle)
{
return;
}
// At a candle close:
// Record as many indicator features as I can think of
// Imagine a trade of both directions was created at the open on this new candle
// See how the trades plays out over the next WindowSize candles
// Output 'buy' 'sell' or 'no trade' based on the trade result
// This means the training data will contain the state of all indicators at this
// candle close, and the ideal trade to have taken at that point.
// Use a position in the past as the 'current' position so that we can
// use the future candle data to see what actually happens.
// 'index - 1' is the candle that just closed, 'index' is the new candle
var index = Instrument.Count - WindowSize;
NegIdx neg_idx = Instrument.FirstIdx + index;
// Get features up to the last closed candle, since we don't know anything about 'index' yet
var features = NNet.Features;
// Create a buy and sell trade at the open of the new candle
// Use the Broker's ChooseSL/TP so that the trades are realistic
var entry = Instrument[neg_idx];
var buy = new Trade(Bot, Instrument, TradeType.Buy, neg_idx: neg_idx);
var sel = new Trade(Bot, Instrument, TradeType.Sell, neg_idx: neg_idx);
// See how each trade plays out over the next WindowSize candles
for (var i = neg_idx; i != Instrument.LastIdx; ++i)
{
var c = Instrument[i];
buy.AddCandle(c, i);
sel.AddCandle(c, i);
}
// Select the best trade as the desired output for the given features
var is_buy = buy.PeakRtR > sel.PeakRtR && buy.Result != Trade.EResult.Open;
var is_sel = sel.PeakRtR > buy.PeakRtR && sel.Result != Trade.EResult.Open;
var labels = "|labels {0} {1} {2} ".Fmt(is_buy?1:0, is_sel?1:0, !(is_buy || is_sel)?1:0);
// Select the file to write to
var fs = (index % TrainToTestRatio) == 0 ? m_testing : m_training;
// Output the labels
fs.Write(labels);
// Write the feature string (last so the training data is easier to read)
fs.Write(features);
// New line
fs.WriteLine();
// Write the results to a file
if (cooldown == 0)
{
if (m_ldr == null)
{
m_ldr = new pr.ldr.LdrBuilder();
m_grp = m_ldr.Group("Training");
}
if (buy.PeakRtR > sel.PeakRtR)
{
Debugging.Dump(buy, ldr_: m_ldr);
}
else
{
Debugging.Dump(sel, ldr_: m_ldr);
}
cooldown = 20;
}
else
{
--cooldown;
}
}
