@eddiee Hi, Mr. Eddie.

I am new to building strategies using ML/DL on Quantiacs and am very impressed with the OS performance of your ML strategies. I hope you can give me your contact (mail, limkedin,...) so I can learn from your experience in building an ML/DL strategy.

Sincerely thank.

@support
Thanks a lot !

@cyan-gloom

Hello. The provided code is insufficient to understand the problem.

I assume that a certain function might not be returning the required value (for instance, the function where your model is being created).

I recommend that you check all return values of functions, using tools like display or print. Then, compare them with what is returned in properly working examples.

The state allows you to use data from previous iterations. You can find an example here:
https://github.com/quantiacs/toolbox/blob/2f4c42e33c7ce789dfad5d170444fd542e28c8ae/qnt/examples/004-strategy-futures-multipass-stateful.py

Thanks for the answer!

I still think something is wrong with this correlation checker. I even used this function to randomize the weights a few times, and I got the same correlation error:

def add_random_noise(weights, noise_level=0.01): noise = np.random.uniform(-noise_level, noise_level, size=weights.shape) return weights + noise

I am pretty sure it's impossible to have 90% correlation in this case.

@magenta-grimer Hello, we updated the documentation.

Now there are 78 futures contracts. Yes, we allow allocating to only 1 asset. If you trade more assets, then you can go long on some of them and short others.

Using more assets helps in increasing the Sharpe ratio, as the mean return grows linearly with the number of assets, and the volatility in the denominator with the square root of the number of assets if there are no correlation terms.

Using uncorrelated assets would then lead to a scaling of the Sharpe ratio with the square root of the number of assets. In practice, however, correlation terms are decreasing this growth.

Stated more simply, it is a good idea to avoid putting all your eggs in the same basket...

@johback

Hello

More examples are here https://github.com/quantiacs/toolbox/blob/main/qnt/tests/test_fundamental_data.py

This is a simple example.

import qnt.data as qndata import datetime as dt import qnt.data.secgov_indicators import qnt.data as qndata import qnt.stats as qns assets = qndata.stocks.load_ndx_list(tail=dt.timedelta(days=5 * 365)) assets_names = [i["id"] for i in assets] data = qndata.stocks.load_ndx_data(tail=dt.timedelta(days=5 * 365), dims=("time", "field", "asset"), assets=assets_names, forward_order=True) facts_names = ['operating_expense'] # 'assets', 'liabilities', 'ivestment_short_term' and other fundamental_data = qnt.data.secgov_load_indicators(assets, time_coord=data.time, standard_indicators=facts_names) # Operating expenses include marketing, noncapitalized R&D, # travel and entertainment, office supply, rent, salary, cogs... weights = fundamental_data.sel(field='operating_expense') is_liquid = data.sel(field="is_liquid") weights = weights * is_liquid # calc stats stats = qns.calc_stat(data, weights.sel(time=slice("2006-01-01", None))) display(stats.to_pandas().tail()) # graph performance = stats.to_pandas()["equity"] import qnt.graph as qngraph qngraph.make_plot_filled(performance.index, performance, name="PnL (Equity)", type="log")

@multi_byte-wildebeest Hi, these limitations refer to the processing time per point in time, not for the full strategy.

If it takes 10 minutes per historical day, and the simulation has to take into account 250 days for let us say 10 years, the multi-pass simulation would process 6 days per hour, 144 days per real day, that means 2 weeks of processing time for the full submission, it is a lot of time.

@support Thx for ur reply

@pink-seel Super that you found it, please do not hesitate to ask for support!

@auxiliary-snail Hi,

unfortunately, this is not allowed and in accordance with the rules. Using hard-coded time periods in which trading algorithm will work differently, is not a quantitative method (just like manual asset selection, e.g. "trade only Apple or Microsoft"). We still haven't implemented a mechanism for automatic recognition of such behaviors in trading strategies, and even though a strategy could be successfully submitted, it will not be eligible for prize winning.
What we are searching for, is well performing strategy over entire in_sample period (SR>0.7), robust to all market movements 2006-2025, so we can expect it will perform well in future, too.

@penrose-moore Thank you for the idea. For the Bitcoin Futures contest we are indeed patching the Bitcoin Futures data with the BTC spot price to build a meaningful time series. For the other Futures contracts, for the moment we will keep the futures histories only, but add spot prices + patching with spot prices to increase the length of the time series to our to-do list.

Thanks !

@support
Cool, thanks very much! 👍

@nosaai Hello

Spyder should be run under conda environment

conda activate qntdev conda install spyder spyder

an alternative way is to clone the library from https://github.com/quantiacs/toolbox
and develop strategies inside qnt. But I recommend using the approach from the documentation.

@multi_byte-wildebeest Hello. I don't use machine learning models in trading.

@dark-pidgeot Yes, of course. Please note that we do not implement leverage, and the sum of the absolute values of the weights has to be equal or smaller than 1. If it is larger, they will be rescaled down.

@support

This is a possible fix, but no gurantee. You have to adjust also the prediction function.

def train_model(data):
"""Create and train the models working on an asset-by-asset basis."""

models = dict()

asset_name_all = data.coords['asset'].values

data = data.sel(time=slice('2013-05-01',None)) # cut the noisy data head before 2013-05-01

features_all = get_features(data)
target_all = get_target_classes(data)

model = create_model()

for asset_name in asset_name_all:

# drop missing values: target_cur = target_all.sel(asset=asset_name).dropna('time', 'any') features_cur = features_all.sel(asset=asset_name).dropna('time', 'any') # align features and targets: target_for_learn_df, feature_for_learn_df = xr.align(target_cur, features_cur, join='inner') if len(features_cur.time) < 10: # not enough points for training continue try: model.fit(feature_for_learn_df.values, target_for_learn_df) models[asset_name] = model except KeyboardInterrupt as e: raise e except: logging.exception('model training failed')

return models

@captain-nidoran Fixed, sorry for issue

@cespadilla Hello.

The reason is in "train_model" function.

def train_model(data): asset_name_all = data.coords['asset'].values features_all = get_features(data) target_all = get_target_classes(data) models = dict() for asset_name in asset_name_all: # drop missing values: target_cur = target_all.sel(asset=asset_name).dropna('time', 'any') features_cur = features_all.sel(asset=asset_name).dropna('time', 'any') target_for_learn_df, feature_for_learn_df = xr.align(target_cur, features_cur, join='inner') if len(features_cur.time) < 10: continue model = get_model() try: model.fit(feature_for_learn_df.values, target_for_learn_df) models[asset_name] = model except: logging.exception('model training failed') return models

If there are less than 10 features for training the model, then the model is not created (if len(features_cur.time) < 10).

This condition makes sense. I would not remove it.

The second thing that can affect is the retraining interval of the model ("retrain_interval").

weights = qnbt.backtest_ml( train=train_model, predict=predict_weights, train_period=2 *365, # the data length for training in calendar days retrain_interval=10 *365, # how often we have to retrain models (calendar days) retrain_interval_after_submit=1, # how often retrain models after submission during evaluation (calendar days) predict_each_day=False, # Is it necessary to call prediction for every day during backtesting? # Set it to true if you suspect that get_features is looking forward. competition_type='crypto_daily_long_short', # competition type lookback_period=365, # how many calendar days are needed by the predict function to generate the output start_date='2014-01-01', # backtest start date analyze = True, build_plots=True # do you need the chart? )

@vyacheslav_b Apologies for the late response. Thanks for the assistance, all is now well. Cheers