-
Notifications
You must be signed in to change notification settings - Fork 7
/
02_var_model.py
344 lines (245 loc) · 11.7 KB
/
02_var_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
# Databricks notebook source
# MAGIC %md
# MAGIC # Model building
# MAGIC In this notebook, we retrieve last 2 years worth of market indicator data to train a model that could predict our instrument returns. As our portfolio is made of 40 equities, we want to train 40 predictive models in parallel, collecting all weights into a single coefficient matrix for monte carlo simulations. We show how to have a more discipline approach to model development by leveraging **MLFlow** capabilities.
# COMMAND ----------
# MAGIC %run ./config/configure_notebook
# COMMAND ----------
import datetime
model_date = datetime.datetime.strptime(config['model']['date'], '%Y-%m-%d')
# COMMAND ----------
# MAGIC %md
# MAGIC We create a temp directory where we may store some additional artifacts for our model
# COMMAND ----------
import tempfile
tempDir = tempfile.TemporaryDirectory()
# COMMAND ----------
# MAGIC %md
# MAGIC ## Compute returns
# MAGIC In the previous notebook, we already computed daily returns for each of our market indicators. Those will be used as features for our model, trying to predict investment returns for each of our instrument.
# COMMAND ----------
from pyspark.sql import functions as F
import pandas as pd
import datetime
market_df = spark.read.table(config['database']['tables']['volatility']).filter(F.col('date') < model_date).select('date', 'features')
market_pd = pd.DataFrame(market_df.toPandas()['features'].to_list(), columns=list(market_indicators.values()))
display(market_pd)
# COMMAND ----------
# MAGIC %md
# MAGIC Let's compute daily returns of our investments. Given the size of a typical portfolio, we can leverage Window functions on spark to do so.
# COMMAND ----------
from pyspark.sql import Window
from pyspark.sql.functions import udf
from pyspark.sql import functions as F
from utils.var_udf import compute_return
def get_stock_returns():
# Apply a tumbling 1 day window on each instrument
window = Window.partitionBy('ticker').orderBy('date').rowsBetween(-1, 0)
# apply sliding window and take first element
stocks_df = spark.table(config['database']['tables']['stocks']) \
.filter(F.col('close').isNotNull()) \
.withColumn("first", F.first('close').over(window)) \
.withColumn("return", compute_return('first', 'close')) \
.select('date', 'ticker', 'return')
return stocks_df
# COMMAND ----------
stocks_df = get_stock_returns().filter(F.col('date') < model_date)
display(stocks_df)
# COMMAND ----------
# MAGIC %md
# MAGIC ## Create features
# MAGIC Risk models are complex and cannot really be expressed simply through the form of notebooks. This solution accelerator does not aim to build best financial model, but rather to walk someone through all processes to do so. The starting point to any good risk model will be to diligently study correlations between all different indicators (limited to 5 here for presentation purpose)
# COMMAND ----------
import seaborn as sns
import matplotlib.pyplot as plt
# we simply plot correlation matrix via pandas (market factors fit in memory)
# we assume market factors are not correlated (NASDAQ and SP500 are, so are OIL and TREASURY BONDS)
f_cor_pdf = market_pd.corr(method='spearman', min_periods=12)
sns.set(rc={'figure.figsize':(11,8)})
sns.heatmap(f_cor_pdf, annot=True)
plt.savefig('{}/factor_correlation.png'.format(tempDir.name))
plt.show()
# COMMAND ----------
# MAGIC %md
# MAGIC We join our market indicator data with stock returns to build an input dataset we can machine learn. We'll use [`tempo`](https://databrickslabs.github.io/tempo/) for this AS-OF join since our timestamps may be different in real life, with intra day tick data.
# COMMAND ----------
from tempo import *
market_tsdf = TSDF(market_df.join(stocks_df.select('ticker').distinct()), ts_col="date", partition_cols=['ticker'])
stocks_tsdf = TSDF(stocks_df, ts_col="date", partition_cols=['ticker'])
# COMMAND ----------
features_df = (
stocks_tsdf.asofJoin(market_tsdf).df
.select(
F.col('date'),
F.col('ticker'),
F.col('right_features').alias('features'),
F.col('return')
)
.filter(F.col('features').isNotNull())
)
display(features_df)
# COMMAND ----------
# MAGIC %md
# MAGIC ## Building models
# MAGIC We show how any function or model can be easily wrapped as a `mlflow.pyfunc` model and registered as such on ml registry. Real life VAR models are obviously more complex than a simple linear regression described here and are not necessarily out of the box sklearn or keras models. Still, they should follow same ML development standard and can easily benefit from ml-flow functionalities as long as one can express model I/O as a form of `pd.Series`, `pd.DataFrame` or `np.array`
# COMMAND ----------
import statsmodels.api as sm
from pyspark.sql.types import *
from pyspark.sql.functions import pandas_udf, PandasUDFType
from utils.var_utils import non_linear_features
# use pandas UDF to train multiple model (one for each instrument) in parallel
# the resulting dataframe will be the linear regression weights for each instrument
train_model_schema = StructType([
StructField('ticker', StringType(), True),
StructField('weights', ArrayType(FloatType()), True)
])
# a model would also be much more complex than the below
@pandas_udf(train_model_schema, PandasUDFType.GROUPED_MAP)
def train_model(group, pdf):
import pandas as pd
import numpy as np
# build market factor vectors
# add a constant - the intercept term for each instrument i.
X = [non_linear_features(row) for row in np.array(pdf['features'])]
X = sm.add_constant(X, prepend=True)
y = np.array(pdf['return'])
model = sm.OLS(y, X).fit()
w_df = pd.DataFrame(data=[[model.params]], columns=['weights'])
w_df['ticker'] = group[0]
return w_df
# COMMAND ----------
# the resulting dataframe easily fits in memory and will be saved as our "uber model"
model_df = features_df.groupBy('ticker').apply(train_model).toPandas()
display(model_df.head(10))
# COMMAND ----------
# MAGIC %md
# MAGIC One can package an entire business logic (being statistical models or more AI models) as a simple `pyfunc`.
# COMMAND ----------
import mlflow
from mlflow.pyfunc import PythonModel
class RiskMLFlowModel(PythonModel):
def __init__(self, model_df):
self.weights = dict(zip(model_df.ticker, model_df.weights))
def _predict_record(self, ticker, xs):
# Our logic is really simplistic and use simple non linear features with a linear regression
# still, models could be packaged as pyfunc regardless of sklearn, complex DL or plain stats objects
from utils.var_utils import non_linear_features
from utils.var_utils import predict_non_linears
ps = self.weights[ticker]
fs = non_linear_features(xs)
return predict_non_linears(ps, fs)
def predict(self, context, model_input):
predicted = model_input[['ticker','features']].apply(lambda x: self._predict_record(*x), axis=1)
return predicted
# COMMAND ----------
# MAGIC %md
# MAGIC Such a model will be tracked, stored, registered and signature of the model enforced to prevent from data drift.
# COMMAND ----------
from mlflow.models.signature import infer_signature
with mlflow.start_run(run_name='value-at-risk') as run:
# get mlflow run ID
run_id = run.info.run_id
# create our pyfunc model
python_model = RiskMLFlowModel(model_df)
# Get model input and output signatures
model_input_df = features_df.select('ticker', 'features').limit(10).toPandas()
model_output_df = python_model.predict(None, model_input_df)
model_signature = infer_signature(model_input_df, model_output_df)
# log our model to mlflow
mlflow.pyfunc.log_model(
artifact_path="model",
python_model=python_model,
signature=model_signature
)
# log additional artifacts
mlflow.log_artifact("{}/factor_correlation.png".format(tempDir.name))
# COMMAND ----------
model_udf = mlflow.pyfunc.spark_udf(model_uri='runs:/{}/model'.format(run_id), result_type='float', spark=spark)
prediction_df = features_df.withColumn('predicted', model_udf(F.struct('ticker', 'features')))
display(prediction_df)
# COMMAND ----------
# compare expected vs. actual return
# sum mean square error per instrument
from utils.var_udf import wsse_udf
wsse_df = prediction_df \
.withColumn('wsse', wsse_udf(F.col('predicted'), F.col('return'))) \
.groupBy('ticker') \
.agg(F.sum('wsse').alias('wsse'))
# get average wsse across portfolio
wsse = wsse_df.select(F.avg('wsse').alias('wsse')).toPandas().iloc[0].wsse
# plot mean square error as accuracy of our model for each instrument
ax = wsse_df.toPandas().plot.bar(x='ticker', y='wsse', rot=0, label=None, figsize=(24,5))
ax.get_legend().remove()
plt.title("Model WSSE for each instrument")
plt.xticks(rotation=45)
plt.ylabel("wsse")
plt.savefig("{}/model_wsse.png".format(tempDir.name))
plt.show()
# COMMAND ----------
# MAGIC %md
# MAGIC We can update our previous experiment with results of our prediction model (sum square of error)
# COMMAND ----------
with mlflow.start_run(run_id=run_id) as run:
mlflow.log_metric("wsse", wsse)
mlflow.log_artifact("{}/model_wsse.png".format(tempDir.name))
# COMMAND ----------
# MAGIC %md
# MAGIC The experiment captured now contains all libraries required to run in isolation and is linked to specific delta version to enable full reproducibility
# COMMAND ----------
# MAGIC %md
# MAGIC <img src=https://d1r5llqwmkrl74.cloudfront.net/notebooks/fsi/var/images/var_experiments.png width="1000px">
# COMMAND ----------
# MAGIC %md
# MAGIC By registering our model to ML registry, we make it available to downstream processes and backend jobs such as our next notebook focused on monte carlo simulations
# COMMAND ----------
client = mlflow.tracking.MlflowClient()
model_uri = "runs:/{}/model".format(run_id)
result = mlflow.register_model(model_uri, config['model']['name'])
version = result.version
# COMMAND ----------
# MAGIC %md
# MAGIC We can also promote our model to different stages programmatically. Although our models would need to be reviewed in real life scenario, we make it available as a production artifact for our next notebook and programmatically transition previous runs back to Archive.
# COMMAND ----------
client = mlflow.tracking.MlflowClient()
for model in client.search_model_versions("name='{}'".format(config['model']['name'])):
if model.current_stage == 'Production':
print("Archiving model version {}".format(model.version))
client.transition_model_version_stage(
name=config['model']['name'],
version=int(model.version),
stage="Archived"
)
# COMMAND ----------
client = mlflow.tracking.MlflowClient()
client.transition_model_version_stage(
name=config['model']['name'],
version=version,
stage="Production"
)
# COMMAND ----------
# MAGIC %md
# MAGIC Our model now production candidate, we can load our predictive logic as a simple user defined function and predict investment returns for every observed market condition
# COMMAND ----------
model_udf = mlflow.pyfunc.spark_udf(
model_uri='models:/{}/production'.format(config['model']['name']),
result_type='float',
spark=spark
)
# COMMAND ----------
import numpy as np
plt.figure(figsize=(25,12))
prediction_df = features_df.withColumn('predicted', model_udf(F.struct("ticker", "features")))
df_past_1 = prediction_df.filter(F.col('ticker') == "EC").orderBy('date').toPandas()
df_past_2 = prediction_df.filter(F.col('ticker') == "EC").orderBy('date').toPandas()
plt.plot(df_past_1.date, df_past_1['return'])
plt.plot(df_past_2.date, df_past_2['predicted'], color='green', linestyle='--')
min_return = np.min(df_past_2['return'])
max_return = np.max(df_past_2['return'])
plt.ylim([min_return, max_return])
plt.title('Log return of EC')
plt.ylabel('log return')
plt.xlabel('date')
plt.show()
# COMMAND ----------
tempDir.cleanup()
# COMMAND ----------