Why ML → Finance is Hard (1/4)

I have used machine learning in trading strategies over the past 10 years or so. However my use of ML has often played a relatively small role in the overall design and success of the strategies due to issues particular to financial data sets. I tend to use ML in specific signals or strategy sub-problems where the data / problem setup have attributes that lead to a robust statistical solution. This is as opposed to the “Nirvana” scenario where fundamental features and objective are provided to an AI and the AI generates functioning strategies with little effort on the part of the researcher.

I was part of a very well funded AI startup that used machine learning on a massive scale to discover new and novel medium frequency trading strategies. More often than not, a strategy determined in training failed to generalize and behave as expected in testing, validation, or live. We spent considerable time and research understanding and improving on the problems of applying ML to trading over the years. Subsequent to this have used machine learning in prop trading, market making, and in derivatives modeling at other firms.

As have had a few learnings from these experiences, hopefully some that are useful to others, will share here:

Problems

Financial timeseries present some of the most difficult problems for machine learning. Here I will list a few of the challenges with financial data and why it is so difficult to model with ML.

1. Low signal to noise:
   • label noise
   • features and measures are noisy variables
2. Feature vector non-independence
   • Many machine learning algorithms require that each sample (feature vector / label) be independent of other samples.
   • Many features make use of windows, creating overlap between the current sample and prior samples sharing part of the same lookback period.
   • This can have a devastating effect on training, causing model bias and overfitting.
3. Standard loss functions not geared for trading objectives
   • We have much higher aversion to losses (false-positives) than to opportunity cost (false-negative), whereas ML models try to balance precision and recall.
   • Many supervised machine learning algorithms need to be adjusted towards a suitable trading objective.
4. Sparsity of opportunities (for some strategies):
   • implication: unbalanced data sets for supervised learning
   • very often a trading opportunity is an “outlier”, a much less frequent event in the data. Most machine learning algorithms are geared for balanced data sets.
5. Non-stationarity
   • Prices not stationary and returns are memory-less, need to balance between the two
   • Timescales are not stationary with respect to patterns (patterns can play out over shorter or longer timescales)
   • Regime change presents shift in fundamental trading patterns -> complexity for training models
6. Lack of data or biased data
   • Depending on frequency of data or pattern one is pursuing, financial data is often orders of magnitudes less plentiful than data sets in other fields.
   • Markets can be in 1 regime for an extended period of time (witness the equity bull markets), biasing available data towards a single regime.

There are ways to deal with some of these issues. I first want to highlight the problems, illustrating with examples.

Toy Example

Suppose we want to use ML to create a long-only model to predict whether the 5 day return > some minimum return (say 50bps).

1. Create labels, identifying trading opportunities
   • We will label 5 day returns >= 50bps with 1 (where we should enter) and < 50bps with 0 (where we should skip entry).
2. Create features for the model
   • Note that I don’t advocate using technical indicators generally, and the features below are just for illustration purposes, not chosen intentionally.
3. Train the model
4. Evaluate

See full code in the appendix

bars = getOHLC ("SPY")
close = bars["Adj Close"].values.flatten()

# create features
df = bars.copy()
df["rsi"] = talib.RSI(close, timeperiod=14)
df["roc5"] = talib.ROC(close, timeperiod=5)
...

# create { 0, 1 } labels, where 1 means 5 day return >= 50bps
df["label"] = (df["roc5"].shift(-5) >= 0.50) * 1.0

# feature columns
features = ["rsi", "roc1", "roc5", "roc10", "roc20", "oc", "hl"]

# split data set
icut = int(df.shape[0] * 0.70)
training = df.iloc[:icut].dropna()
testing = df.iloc[icut:].dropna()

# train model on features & labels
clf = RandomForestClassifier(n_estimators=500, random_state=1, n_jobs=-1)
model = clf.fit (training[features], training.label)

# use model to predict labels for training period and testing period respectively
pred_train = model.predict(training[features])
pred_test = model.predict(testing[features])

Now lets look at the confusion matrix for training and testing through the model we trained. For trading we want to maximize TP (true-positive, our profitable trades) and minimize FP (false positives, out losing trades):

For trading we are less concerned with TN (true negative) and FN (false negative). False negatives would represent missed opportunities, but no loss.

Now let’s evaluate the confusion matrices for the model given the training data and then the testing data. Our hope is that the testing (out-of-sample) data presents an accuracy similar to training and that the ratio of TP to FP is high.

confusion_matrix(training.label, pred_train)
confusion_matrix(testing.label, pred_test)

It turns out that our training period has a perfect fit (FN = 0 and FP = 0). This is a sure sign of overfitting:

And in testing (out-of-sample) has a poor precision where 271 of (271 + 241) trades are losing.

Discussion

The above example illustrates a number of problems we will discuss in subsequent posts:

• Labeling is noisy (noisy labeling problem)
  • We have assigned +1 labels to returns >= 50bps, however some of these returns may not be representative of an underlying move, rather be a presentation of noise around the true price. For example, the true price return over a period could be 0, but have a variance of 50bps or more.
  • We have assigned 0 labels to returns < 50bps, however in some cases the underlying price process may be yielding >= 50bps over the period, but the sample return is reduced below 50bps due to volatility around a 50bps+ mean.
  • Hence we are pushing the model to map features to noise in some % of instances and biasing the model
• Features are noisy
  • Our open-close and high-low features, in particular, will suffer from volatility
• Samples not independent
  • Each row in our feature set is not independent from neighboring rows. Our longest feature has a lookback window of 20 bars. Hence every feature row will have overlap with 40 other features (20 in the past + 20 in the future). Many ML algorithms will exploit the information leakage due to non-independence, creating an overfit model in training.
  • There are changes we can make to RandomForest, for example, which will help us overcome this problem.
• Biased data set
  • The data set used in this example is primarily showing prices rising over time (due to the “unnatural” and continual rise of the SPY over the last 10+ years). It would be better to have data which expresses equal weighted upward and downward trends.
• Wrong loss objective
  • We would prefer to optimize for a better balance between TP and FP (precision), as opposed to balancing precision and recall.

Will discuss these in more depth in subsequent posts, and pose some adjustments that can be done to algorithms, datasets, etc.

See Next

• part 2
• part 3

Appendix

import talib
import pandas as pd
import numpy as np
from datetime import datetime
from pandas_datareader import data as pdr
from sklearn.metrics import confusion_matrix
from sklearn.ensemble import RandomForestClassifier

def getOHLC (stock: str, Tstart = datetime(1999,1,1), Tend = datetime.now()):
    raw = pdr.get_data_yahoo([stock], start=Tstart, end=Tend)
    df = raw[["Open","High","Low","Close","Adj Close","Volume"]]
    return df

bars = getOHLC ("SPY")
close = bars["Adj Close"].values.flatten()

# create features
df = bars.copy()
df["rsi"] = talib.RSI(close, timeperiod=14)
df["roc1"] = talib.ROC(close, timeperiod=1)
df["roc5"] = talib.ROC(close, timeperiod=5)
df["roc10"] = talib.ROC(close, timeperiod=10)
df["roc20"] = talib.ROC(close, timeperiod=20)
df["oc"] = np.log(bars.Close / bars.Open)
df["hl"] = np.log(bars.High / bars.Low)

# create { 0, 1 } labels, where 1 means 5 day return >= 50bps
df["label"] = (df["roc5"].shift(-5) >= 0.50) * 1.0

# feature columns
features = ["rsi", "roc1", "roc5", "roc10", "roc20", "oc", "hl"]

# split data set
icut = int(df.shape[0] * 0.70)
training = df.iloc[:icut].dropna()
testing = df.iloc[icut:].dropna()

# train model on features & labels
clf = RandomForestClassifier(n_estimators=500, random_state=1, n_jobs=-1)
model = clf.fit (training[features], training.label)

# use model to predict labels for training period and testing period respectively
pred_train = model.predict(training[features])
pred_test = model.predict(testing[features])

print(confusion_matrix(training.label, pred_train))
print(confusion_matrix(testing.label, pred_test))

# note that this is an overestimate, since may double count overlapping +1 label 5 day periods
print("P&L estimate: %1.0f%%" % (pred_test * testing.rfwd).sum())