This NN model performs worse than sklearn's `LogisticRegression` model (with default params too). Really I can only get about 57% on the test set, and LR gives me about 58%. What am I doing wrong - I've tried dozen of permutations of the network topology, optimisers, etc.? Note the size of the data is (400k, 144), so I have lots of data.

model.py

# FFN to model game outcome
import pandas as pd
from keras.layers import Dense
from keras.models import Sequential
import numpy as np


if __name__ == "__main__":
    import sqlite3
    with sqlite3.connect("data/heroes.db") as conn:
        with open("data/sql/feature_matrix.sql", "r") as sql:
            df = pd.read_sql(sql.read(), conn)
            df = df.drop("ReplayID", axis=1)
            df = df.dropna(axis=0, how='any')

    TRAINING_DATA = 400000
    # shape = (TRAINING_DATA, 144)
    X_train = df.iloc[:TRAINING_DATA].drop('outcome', axis=1)
    Y_train = df.iloc[:TRAINING_DATA]['outcome']

    X_test = df.iloc[TRAINING_DATA:].drop('outcome', axis=1)
    Y_test = df.iloc[TRAINING_DATA:]['outcome']

    model = Sequential()
    model.add(Dense(100, activation='relu', kernel_initializer='he_normal', input_dim=X_train.shape[1]))
    model.add(Dense(10, activation='relu', kernel_initializer='he_normal'))
    model.add(Dense(1, activation='sigmoid'))
    model.compile(optimizer='rmsprop',
                  loss='binary_crossentropy',
                  metrics=['binary_accuracy'])
    model.fit(X_train.values, Y_train.values, epochs=50, batch_size=512)
    print model.evaluate(X_test.values, Y_test.values, batch_size=128)

could you share the data too?

Try shuffling the data

You probably have to standardize your features. Given that sklearn's LogisticRegression trains through coordinate descent by deafult (which is not that sensitive to feature scale), LogisticRegression is working better. Try to use sklearn's StandardScaler before training your network.