def main(stocks, target_brand, layer1, layer2, load_sess, result_file=None):
# 学習データのキャッシュ
feed_cache = FeedCache(target_brand, str(REMOVE_NEWEST_DAYS))
# 対象の銘柄名
target_brand_name = brand_name(target_brand)
if not feed_cache.is_exist():
print('Make cache')
# 株価指標データを読み込む
all_data = load_exchange_dataframes(stocks, target_brand)
# 終値を取得
using_data = get_using_data(all_data, target_brand)
# データを学習に使える形式に正規化
log_return_data = get_log_return_data(stocks, using_data)
# 答と学習データを作る
training_test_data = build_training_data(stocks, log_return_data,
target_brand)
feed_cache.save(training_test_data)
else:
print('Exist cache')
training_test_data = feed_cache.load()
# 学習データをトレーニング用とテスト用に分割する
dataset = split_training_test_data(CLASS_COUNT, training_test_data)
if len(dataset.training_predictors) < TRAIN_MIN:
print('[{}]{}: 学習データが少なすぎるため計算を中止'.format(target_brand,
target_brand_name))
with open('results.csv', 'a') as f:
f.write('{},{},ERROR\n'.format(target_brand, target_brand_name))
# レイヤー検証ログに保存
brands = nikkei225_s
codes = [code for (code, name, _) in brands]
layerLog = LayerLog('layer_logs', '{}_{}.csv'.format(layer1, layer2),
codes)
layerLog.add(target_brand, [-1, 0, 0, 0, 0, 0, 0, 0])
return
print('[{}]{}'.format(target_brand, target_brand_name))
# 器械学習のネットワークを作成
env = smarter_network(stocks, dataset, layer1, layer2)
# 学習
train_accuracy, money, trues, falses, actual_count, deal_logs, up_expectation_dates = train(
load_sess, env, stocks[target_brand], target_brand)
print('-- テスト --')
# 各クラスの正解率
rates = np.zeros(CLASS_COUNT, dtype=np.int64)
# 各クラスの正解数
counts = np.zeros(CLASS_COUNT, dtype=np.int64)
for i in range(CLASS_COUNT):
counts[i] = trues[i] + falses[i]
if counts[i]:
# 各クラスの正解率（予想数 / 正解数）
rates[i] = int(float(trues[i]) / float(counts[i]) * 100)
print('下げ正解率 : {}% 予想{}回'.format(rates[CLASS_DOWN], counts[CLASS_DOWN]))
print('変化なし正解率: {}% 予想{}回'.format(rates[CLASS_NEUTRAL],
counts[CLASS_NEUTRAL]))
print('上げ正解率 : {}% 予想{}回'.format(rates[CLASS_UP], counts[CLASS_UP]))
print('上げ予測日 : {}'.format(up_expectation_dates))
print('-- 売買シミュレーション --')
print('売買シミュレーション結果 {:,d}円'.format(money))
# 結果をファイル保存
if result_file:
with open(result_file, 'a') as f:
f.write('{},{},{},{},{},{},{},{},{}\n'.format(
target_brand, target_brand_name, money, trues[CLASS_DOWN],
falses[CLASS_DOWN], trues[CLASS_NEUTRAL],
falses[CLASS_NEUTRAL], trues[CLASS_UP], falses[CLASS_UP]))
else:
print(target_brand, target_brand_name, money, trues[CLASS_DOWN],
falses[CLASS_DOWN], trues[CLASS_NEUTRAL], falses[CLASS_NEUTRAL],
trues[CLASS_UP], falses[CLASS_UP])
# 売買履歴をファイルに保存
save_deal_logs(target_brand, deal_logs)
# レイヤー検証ログに保存
brands = nikkei225_s
codes = [code for (code, name, _) in brands]
layerLog = LayerLog('layer_logs', '{}_{}.csv'.format(layer1, layer2),
codes)
layerLog.add(target_brand, [
train_accuracy, money, trues[CLASS_DOWN], falses[CLASS_DOWN],
trues[CLASS_NEUTRAL], falses[CLASS_NEUTRAL], trues[CLASS_UP],
falses[CLASS_UP]
])
# 購入予想日を保存する
save_up_expectation_dates(target_brand, up_expectation_dates)
#print('器械学習のネットワークを作成')
env = smarter_network(dataset)
if args.inspect:
import code
print('Press Ctrl-d to proceed')
code.interact(local=locals())
#print('学習')
return train(env, steps=args.steps, checkin_interval=args.checkin)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('target_brand')
parser.add_argument('--steps', type=int, default=10000)
parser.add_argument('--checkin', type=int, default=1000)
parser.add_argument('--use-subset', type=float, default=None)
parser.add_argument('--inspect', default=False, action='store_true')
args = parser.parse_args()
stocks[args.target_brand] = Stock(YahooJp, args.target_brand)
result = main(args)
target_brand_name = brand_name(args.target_brand)
print('[{}]{}: {}'.format(args.target_brand, target_brand_name, result['Accuracy']))
with open('results.csv', 'a') as f:
f.write('{},{},{}\n'.format(args.target_brand, target_brand_name, str(result['Accuracy'])))
env = smarter_network(dataset)
if args.inspect:
import code
print('Press Ctrl-d to proceed')
code.interact(local=locals())
#print('学習')
return train(env, steps=args.steps, checkin_interval=args.checkin)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('target_brand')
parser.add_argument('--steps', type=int, default=10000)
parser.add_argument('--checkin', type=int, default=1000)
parser.add_argument('--use-subset', type=float, default=None)
parser.add_argument('--inspect', default=False, action='store_true')
args = parser.parse_args()
stocks[args.target_brand] = Stock(YahooJp, args.target_brand)
result = main(args)
target_brand_name = brand_name(args.target_brand)
print('[{}]{}: {}'.format(args.target_brand, target_brand_name,
result['Accuracy']))
with open('results.csv', 'a') as f:
f.write('{},{},{}\n'.format(args.target_brand, target_brand_name,
str(result['Accuracy'])))
def main(stocks, target_brand, layer1, layer2, load_sess, result_file=None):
# 学習データのキャッシュ
feed_cache = FeedCache(target_brand, str(REMOVE_NEWEST_DAYS))
# 対象の銘柄名
target_brand_name = brand_name(target_brand)
if not feed_cache.is_exist():
print('Make cache')
# 株価指標データを読み込む
all_data = load_exchange_dataframes(stocks, target_brand)
# 終値を取得
using_data = get_using_data(all_data, target_brand)
# データを学習に使える形式に正規化
log_return_data = get_log_return_data(stocks, using_data)
# 答と学習データを作る
training_test_data = build_training_data(
stocks, log_return_data, target_brand)
feed_cache.save(training_test_data)
else:
print('Exist cache')
training_test_data = feed_cache.load()
# 学習データをトレーニング用とテスト用に分割する
dataset = split_training_test_data(CLASS_COUNT, training_test_data)
if len(dataset.training_predictors) < TRAIN_MIN:
print('[{}]{}: 学習データが少なすぎるため計算を中止'.format(target_brand, target_brand_name))
with open('results.csv', 'a') as f:
f.write('{},{},ERROR\n'.format(target_brand, target_brand_name))
# レイヤー検証ログに保存
brands = nikkei225_s
codes = [code for (code, name, _) in brands]
layerLog = LayerLog('layer_logs', '{}_{}.csv'.format(layer1, layer2), codes)
layerLog.add(
target_brand,
[-1, 0, 0, 0, 0, 0, 0, 0]
)
return
print('[{}]{}'.format(target_brand, target_brand_name))
# 器械学習のネットワークを作成
env = smarter_network(stocks, dataset, layer1, layer2)
# 学習
train_accuracy, money, trues, falses, actual_count, deal_logs, up_expectation_dates = train(load_sess, env, stocks[target_brand], target_brand)
print('-- テスト --')
# 各クラスの正解率
rates = np.zeros(CLASS_COUNT, dtype=np.int64)
# 各クラスの正解数
counts = np.zeros(CLASS_COUNT, dtype=np.int64)
for i in range(CLASS_COUNT):
counts[i] = trues[i] + falses[i]
if counts[i]:
# 各クラスの正解率（予想数 / 正解数）
rates[i] = int(float(trues[i]) / float(counts[i]) * 100)
print('下げ正解率 : {}% 予想{}回'.format(rates[CLASS_DOWN], counts[CLASS_DOWN]))
print('変化なし正解率: {}% 予想{}回'.format(rates[CLASS_NEUTRAL], counts[CLASS_NEUTRAL]))
print('上げ正解率 : {}% 予想{}回'.format(rates[CLASS_UP], counts[CLASS_UP]))
print('上げ予測日 : {}'.format(up_expectation_dates))
print('-- 売買シミュレーション --')
print('売買シミュレーション結果 {:,d}円'.format(money))
# 結果をファイル保存
if result_file:
with open(result_file, 'a') as f:
f.write('{},{},{},{},{},{},{},{},{}\n'.format(
target_brand,
target_brand_name,
money,
trues[CLASS_DOWN], falses[CLASS_DOWN],
trues[CLASS_NEUTRAL], falses[CLASS_NEUTRAL],
trues[CLASS_UP], falses[CLASS_UP]))
else:
print(
target_brand,
target_brand_name,
money,
trues[CLASS_DOWN], falses[CLASS_DOWN],
trues[CLASS_NEUTRAL], falses[CLASS_NEUTRAL],
trues[CLASS_UP], falses[CLASS_UP])
# 売買履歴をファイルに保存
save_deal_logs(target_brand, deal_logs)
# レイヤー検証ログに保存
brands = nikkei225_s
codes = [code for (code, name, _) in brands]
layerLog = LayerLog('layer_logs', '{}_{}.csv'.format(layer1, layer2), codes)
layerLog.add(
target_brand,
[
train_accuracy,
money,
trues[CLASS_DOWN], falses[CLASS_DOWN],
trues[CLASS_NEUTRAL], falses[CLASS_NEUTRAL],
trues[CLASS_UP], falses[CLASS_UP]
]
)
# 購入予想日を保存する
save_up_expectation_dates(target_brand, up_expectation_dates)