Free lesson · Financial machine learning

Choose the model’s job: targets, horizons and decision layers

Open interactive lesson Practice calculations Explore labs

Start with the idea

Machine learning estimates a relationship from examples. In trading, the most useful target is the one that improves a defined decision—not necessarily tomorrow’s raw closing price.

Symbols, units & horizon

X_t: feature vector genuinely available at decision time
θ: fitted model parameters
f_θ: prediction function
y_(t,H): future simple return label over horizon H
P_entry,P_exit: consistently specified executable or reference prices
μ̂: predicted conditional return in the same units as the label
H: bars or time interval, explicitly fixed

When and why to use this

Use target design before architecture selection. It determines what the learner can improve and which execution assumptions enter its apparent accuracy.

A supervised example contains features available at decision time and a label observed later. Regression can estimate forward residual return, volatility, implementation shortfall or expected loss. Classification can estimate convergence, a positive net trade outcome or a liquidity event. Ranking can order assets for a constrained selection process.

Separate an alpha model from a risk model and an execution model. Alpha predicts conditional opportunity; risk predicts uncertainty or joint losses; execution predicts cost and fill probability. Combining them in a decision layer makes it easier to identify which component adds value and which assumptions fail.

For a forward-return target, use an executable start price and a horizon consistent with the strategy. Barrier labels answer a path-dependent question, but their stop times overlap and require interval-aware purging. A binary sign discards payoff magnitude; a probability above .5 alone need not justify a trade.

y_{t, H} = \frac{P_{t + H}^{e x i t}}{P_{t}^{e n t r y}} - 1, {\hat{μ}}_{t} = f_{θ} (X_{t})

Model assumptions, derivation and arithmetic

Choose the model’s job: targets, horizons and decision layers

Divide the future exit price by the feasible entry price to obtain a wealth ratio. Subtract one for a simple return.
Collect features only from information preceding the decision and retain the label’s availability end time.
Fit a function on training pairs (X,y), then evaluate its output against unseen labels and the strategy’s cost hurdle. The equation defines a supervised task, not a causal economic mechanism.

Work it by hand

An executable entry of 100.10 and exit of 100.40 give label .002997, approximately 29.97 bps. A label based on a midpoint entry of 100 would overstate the executable move.

Apply it in a strategy

Choose one role—alpha, risk, cost, fill or ranking—and state the downstream decision it changes.
Define target units, horizon, label maturity and executable price convention; audit example rows by hand.
Evaluate against a zero, mean, linear or fixed-policy baseline appropriate to that role.

Research deliverable

Create a target contract and feature/label timeline for one real strategy component, including the decision that uses its prediction.

Research sources, review dates and limitations

Python implementation

Self-contained teaching example. Python 3.10+; dependencies and input conventions are shown in the code and notation. Run in your own Python environment.

def forward_return(entry,exit):
    if entry<=0 or exit<=0: raise ValueError("Positive prices required")
    return exit/entry-1

def linear_forecast(features,weights,intercept=0):
    if len(features)!=len(weights): raise ValueError("Dimensions must match")
    return intercept+sum(x*w for x,w in zip(features,weights))

print(forward_return(100.10,100.40))

Continue learning

Machine Learning for Quantitative Strategy Development — all lessons

Quantitative finance and development glossary · Python resources and libraries · Research sources and limitations