MODWT-MARS.py

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from alpha_vantage.timeseries import TimeSeries
from pyearth import Earth
from sklearn.ensemble import AdaBoostRegressor
from sklearn.metrics import mean_squared_error

np.random.seed(55)

# Load MRA detail and scale coefficients
mra = pd.read_csv("modwt_mra.csv")

d1, d2, d3, s3 = mra["d1"], mra["d2"], mra["d3"], mra["s3"]

# Get stock data over same period as MRA
with open("key.txt") as f:
    key = f.readline().strip("\n")

ts = TimeSeries(key=key, output_format="pandas")
data, meta_data = ts.get_daily_adjusted("SPY", outputsize="full")
data = data.loc[data.index >= "2009-01-02"]

price = data["4. close"].iloc[::-1]
returns = np.log(price).diff().dropna()


# Get price labels using same AR scheme
series = returns

D = 4  # number of regressors
T = 1  # delay
N = len(series)
data = np.zeros((N-T-(D-1)*T, D))
lbls = np.zeros((N-T-(D-1)*T,))

for t in range((D-1)*T, N-T):
    data[t-(D-1)*T, :] = [series[t-3*T], series[t-2*T], series[t-T], series[t]]
    lbls[t-(D-1)*T] = series[t+T]
trnLbls = lbls[:lbls.size - round(lbls.size*0.3)]
chkLbls = lbls[lbls.size - round(lbls.size*0.3):]


def MODWT_MARS(series, regressors=4, delay=1):
    src = series.values

    D = regressors  # number of regressors
    T = delay  # delay
    N = len(src)
    data = np.zeros((N-T-(D-1)*T, D))
    lbls = np.zeros((N-T-(D-1)*T, ))

    for t in range((D-1)*T, N-T):
        data[t-(D-1)*T, :] = [src[t-3*T], src[t-2*T], src[t-T], src[t]]
        lbls[t-(D-1)*T] = src[t+T]

    trnData = data[:lbls.size - round(lbls.size*0.3), :]
    trnLbls = lbls[:lbls.size - round(lbls.size*0.3)]
    chkData = data[lbls.size - round(lbls.size*0.3):, :]
    chkLbls = lbls[lbls.size - round(lbls.size*0.3):]

    # append last 4 labels to predict
    a = np.array(chkLbls[-4:]).reshape(1, -1)
    chkData = np.append(chkData, a, axis=0)

    mars = Earth()
    mars.fit(trnData, trnLbls)

    boosted_mars = AdaBoostRegressor(estimator=mars,
                                     n_estimators=50,
                                     learning_rate=0.1,
                                     loss="exponential")

    boosted_mars.fit(trnData, trnLbls)
    oos_preds = boosted_mars.predict(chkData)

    return oos_preds


# MRA
D1_test = pd.DataFrame(MODWT_MARS(d1), columns=["D1"])
D2_test = pd.DataFrame(MODWT_MARS(d2), columns=["D2"])
D3_test = pd.DataFrame(MODWT_MARS(d3), columns=["D3"])
S3_test = pd.DataFrame(MODWT_MARS(s3), columns=["S3"])

pred = pd.concat([D1_test, D2_test, D3_test, S3_test], axis=1)

pred["sum"] = pred.sum(axis=1)


def mda(t: np.ndarray, p: np.ndarray):
    """ Mean Directional Accuracy """
    return np.mean((np.sign(t[1:]-t[:-1]) == np.sign(p[1:]-t[:-1])))


# Metrics (RMSE and DA were used in the paper)
mda = mda(chkLbls, pred["sum"][:-1])
rmse = mean_squared_error(chkLbls, pred["sum"][:-1], squared=False)

# Plot additive prediction against actual returns
plt.plot(np.array(pred["sum"]), color="g")
plt.plot(chkLbls)

y = max(chkLbls)
plt.annotate("Mean Directional Accuracy = {:.3f}".format(mda), (-20, y+0.01))
plt.annotate("RMSE = {:.3e}".format(rmse), (-20, y))
plt.title("SPY Returns prediction using MODWT-MARS Framework")
plt.legend(["Pred", "Actual"])

figure = plt.gcf()
figure.set_size_inches(8, 6)

plt.savefig("Results/test.png", dpi=252)
plt.show()