def train():
# We train different models to predict different outputs.
# 0 = 15th percentile, 1 = 25th percentile, 2 = 35th percentile,
# 3 = 50th percentile, 4 = 65th percentile, 5 = 75th percentile,
# 6 = 85th percentile
outputToPredict = 0
# This represents if we are predicting a binary output (e.g. tomorrow's 50th
# percentile > today's mean) or if we are predicting a percentage (e.g.
# tomorrow's 50th percentile is 97% of today's mean)
binary = BINARY_PREDICTION
print("Loading dataset...")
datasetLoader = DatasetLoader()
# For our outputs, only put in the 15th percentile as the output.
# Note: THIS IS ONLY USING THE TOP INDICATORS!
data, labels = datasetLoader.load(
shuffle=False,
path="../../data_set/final-train-dataset.csv",
onlyLabelToUse=outputToPredict,
useOnlyBestIndicators=False,
binary=binary)
# Hyperparameters!
# learningRate = 0.00005
learningRate = 0.0002
epochs = 200
batchSize = 64
dropout = 0.01
# Not currently in use:
# decayRate = 0.03
# decayStep = 1.0
model = AttentionModel(tryUsingGPU=True, binary=binary)
model.setup(Hyperparameters(learningRate, epochs, dropout, batchSize))
# If we want to predict all the percentiles at once:
# model.createModel(OUTPUT_CHANNELS, generateGraph=False)
# If we only want to predict one of the percentiles:
model.createModel(generateGraph=False)
# model.load()
epochs, hist = model.trainModel(data, labels, 0.15)
listOfMetricsToPlot = model.listOfMetrics
model.plotCurve(epochs, hist)
model.save()
# Test the model on test data.
testData, testLabels = datasetLoader.load(
shuffle=False,
path="../../data_set/final-test-dataset.csv",
onlyLabelToUse=outputToPredict,
useOnlyBestIndicators=False)
model.evaluate(testData, testLabels)
print(model.predict(testData[50:51, :, :]))
print(testData.shape[0])
def testPerformance():
# We are going to need data from 15 days ago up to and including yesterday.
# Make sure to download all the data using data_downloads/DownloadBinanceData.py!
# now = datetime.now()
now = datetime(year=2021, month=3, day=3, hour=23, minute=35)
timezone = "Etc/GMT-0"
timezone = pytz.timezone(timezone)
now = timezone.localize(now)
now = now.replace(hour=0, minute=0, second=0, microsecond=0)
startDate = now - timedelta(days=193)
listOfStocks = ["BTCUSDT"]
historicalObtainer = HistoricalDataObtainer(startDate, now,
"../data_downloads/")
print("Reading historical stock data...")
historicalObtainer.trackStocks(listOfStocks)
model = CnnRnnMlpModel(tryUsingGPU=False)
# Hyperparameters!
learningRate = 0.003
epochs = 500
batchSize = 40
decayRate = 0.005
decayStep = 1.0
dropout = 0.1
model.setup(
Hyperparameters(learningRate,
epochs,
dropout,
batchSize,
decayRate=decayRate,
decayStep=decayStep))
model.createModel()
model.load()
inputData = prepareData(listOfStocks[0], historicalObtainer, startDate,
now)
inputData = np.array(inputData[-1]).T
df = historicalObtainer.getHistoricalDataAsDataframe(listOfStocks[0])
startIndex = df.index[df["Timestamp"] == now - timedelta(days=15)].tolist()
assert len(startIndex) != 0
startIndex = startIndex[0]
endIndex = df.index[df["Timestamp"] == now].tolist()
assert len(endIndex) != 0
endIndex = endIndex[0]
meanPrice = df["Close"].iloc[startIndex:endIndex].mean()
predictions = model.makePricePredictionForTommorrow(inputData, meanPrice)
print("Predictions: [min, 25th percentile, median, 75th percentile, max]")
print(predictions)
def testPerformance():
datasetLoader = DatasetLoader()
trainData, trainLabels = datasetLoader.load(path="../data_set/final-train-dataset.csv", shuffle=False)
testData, testLabels = datasetLoader.load(path="../data_set/final-test-dataset.csv", shuffle=False)
model = CnnRnnMlpModel(tryUsingGPU=True)
# Hyperparameters!
learningRate = 0.003
epochs = 500
batchSize = 40
decayRate = 0.005
decayStep = 1.0
dropout = 0.1
model.setup(Hyperparameters(learningRate, epochs, dropout, batchSize,
decayRate=decayRate, decayStep=decayStep))
model.createModel()
model.load()
predictions = model.predict(trainData, concatenate=True)
reporter = PerformanceReporter()
reporter.reportOnPoorPredictions(predictions, trainLabels, errorForPoor=0.01)
print("The model's performance on the test set.")
reporter.reportPerformanceOnDataset(model, testData, testLabels)