⏳ Quick Start

Use application model

There are several pre-defined fitted models at ml_investment.applications. It incapsulating data and weights downloading, pipeline creation and model fitting. So you can just use it without knowing internal structure.

from ml_investment.applications.fair_marketcap_yahoo import FairMarketcapYahoo

fair_marketcap_yahoo = FairMarketcapYahoo()
fair_marketcap_yahoo.execute(['AAPL', 'FB', 'MSFT'])

ticker	date	fair_marketcap_yahoo
AAPL	2020-12-31	5.173328e+11
FB	2020-12-31	8.442045e+11
MSFT	2020-12-31	4.501329e+11

Create your own pipeline

1. Download data

You may download default datasets by ml_investment.download_scripts

from ml_investment.download_scripts import download_yahoo
from ml_investment.utils import load_config

# Config located at ~/.ml_investment/config.json
config = load_config()

download_yahoo.main(config['yahoo_data_path'])

>>> 1365it [03:32,  6.42it/s]
>>> 1365it [01:49,  12.51it/s]

2. Create dict with dataloaders

You may choose from default ml_investment.data_loaders or wrote your own. Each dataloader should have load(index) interface.

from ml_investment.data_loaders.yahoo import YahooQuarterlyData, YahooBaseData

data = {}
data['quarterly'] = YahooQuarterlyData(config['yahoo_data_path'])
data['base'] = YahooBaseData(config['yahoo_data_path'])

3. Define and fit pipeline

You may specify all steps of pipeline creation. Base pipeline consist of the folowing steps:

Create data dict(it was done in previous step)
Define features. Features is a number of values and characteristics that will be calculated for model trainig. Default feature calculators are located at ml_investment.features
Define targets. Target is a final goal of the pipeline, it should represent some desired useful property. Default target calculators are located at ml_investment.targets
Choose model. Model is machine learning algorithm, core of the pipeline. It also may incapsulate validation and other stuff. You may use wrappers from ml_investment.models

import lightgbm as lgbm
from ml_investment.utils import load_config, load_tickers
from ml_investment.features import QuarterlyFeatures, BaseCompanyFeatures,\
                                   FeatureMerger
from ml_investment.targets import BaseInfoTarget
from ml_investment.models import LogExpModel, GroupedOOFModel
from ml_investment.pipelines import Pipeline
from ml_investment.metrics import median_absolute_relative_error

fc1 = QuarterlyFeatures(data_key='quarterly',
                        columns=['netIncome',
                                 'cash',
                                 'totalAssets',
                                 'ebit'],
                        quarter_counts=[2, 4, 10],
                        max_back_quarter=1)

fc2 = BaseCompanyFeatures(data_key='base', cat_columns=['sector'])

feature = FeatureMerger(fc1, fc2, on='ticker')

target = BaseInfoTarget(data_key='base', col='enterpriseValue')

base_model = LogExpModel(lgbm.sklearn.LGBMRegressor())
model = GroupedOOFModel(base_model=base_model,
                        group_column='ticker',
                        fold_cnt=4)

pipeline = Pipeline(data=data,
                    feature=feature,
                    target=target,
                    model=model,
                    out_name='my_super_model')

tickers = load_tickers()['base_us_stocks']
pipeline.fit(tickers, metric=median_absolute_relative_error)

>>> {'metric_my_super_model': 0.40599471294301914}

4. Inference your pipeline

Since ml_investment.models.GroupedOOFModel was used, there are no data leakage and you may use pipeline on the same company tickers.

pipeline.execute(['AAPL', 'FB', 'MSFT'])

ticker	date	my_super_model
AAPL	2020-12-31	8.170051e+11
FB	2020-12-31	3.898840e+11
MSFT	2020-12-31	3.540126e+11