Trading Models: Predict Direction or Price?

“Will Bitcoin go up?” sounds like the natural question for a trading model. It is often the wrong final question.

A market can rise exactly as predicted and still produce a losing trade if the move is smaller than spread, fees, slippage or funding. A model can also be wrong about the most likely direction while identifying an attractive asymmetric payoff. The useful target depends on the decision the system must make.

Classification compresses the problem

A direction classifier might output:

1 if the price is higher in one hour; and
0 if the price is lower or unchanged.

That framing is easy to score with accuracy. But it throws away magnitude. A one-cent increase and a $5,000 increase receive the same label. A one-cent decline and a liquidation event also share a label.

It can also produce misleading class balance. If an asset rises in 52% of sampled periods, a model that always predicts “up” achieves 52% accuracy without creating a tradable edge.

Regression asks how much

A regression model might estimate a future return, price change or fair value. That output can be compared with the executable price and costs.

Suppose a model estimates BTC’s one-hour fair value at $100,080 while the current best ask is $100,000. Expected gross upside is $80. If estimated round-trip spread, fees and slippage total $120, the correct action may still be do nothing.

The model output is not the order. A decision layer combines it with:

current bid and ask;
uncertainty around the estimate;
position and inventory;
order size and expected impact;
transaction and financing costs;
holding period; and
risk limits.

Probability needs payoff

A calibrated probability can be more useful than a hard label, especially for prediction markets and discrete events. But probability alone still does not determine value.

If a contract trades at $0.70 and a model estimates a 72% chance of a $1 payout, the apparent two-cent edge must cover fees, spread, uncertainty, settlement and resolution risk. If the estimate is poorly calibrated or the contract wording is misunderstood, the calculation is meaningless.

Our Polymarket experiment design uses time-locked forecasts and proper scoring rules so forecast quality is measured separately from simulated execution.

Trading utility is not ordinary model accuracy

Headlands’ review of The Elements of Statistical Learning argues that many trading decisions do not fit an ordinary independent-sample classification frame. Buying, selling, risk, messaging and constraints depend on P&L and on previous state.

That does not mean classification is never useful. It can classify regimes, data quality, event types or the probability that a move exceeds a threshold. The mistake is treating a convenient model metric as the business objective.

Two strategies with identical prediction accuracy can have radically different results:

Strategy	Accuracy	Average win	Average loss	Before costs
A	55%	$1	$1	Positive
B	55%	$1	$3	Negative

Costs, serial dependence and sizing can change both again.

The label changes the research result

Before training, define:

decision time;
prediction horizon;
reference and execution prices;
treatment of unchanged prices;
whether the target includes fees and funding;
overlapping samples;
order and position rules; and
abstention behavior.

If a one-hour target is calculated every minute, neighboring labels overlap heavily. Randomly splitting them can put almost the same market episode in training and test data. Preserve time order and account for overlapping outcomes.

A practical output contract

Instead of “BUY” prose, ask a research model for structured fields:

estimated fair value or return distribution;
horizon and timestamp;
uncertainty interval;
evidence and data version;
known invalidation conditions; and
an explicit abstain state.

Deterministic software should then compare that output with live price, costs, exposure and policy. An LLM’s confidence language is not a risk control.

What can go wrong?

Accuracy is optimized instead of net decision value.
The target price is not the price available to trade.
Overlapping labels leak across random splits.
A model is forced to trade when the edge is smaller than uncertainty.
Magnitude, costs and inventory are ignored.
Probabilities are never checked for calibration.

What this does not prove

Regression is not automatically superior to classification. Simple rule-based models can outperform complex ones when the target and execution are better specified. The correct model form follows the decision, data and validation—not fashion.

Next: Backtest → Historical Replay → Paper Trading: The Validation Ladder.

Risk disclosure: Model outputs are uncertain estimates, not facts. Prediction accuracy does not guarantee executable profit, and automation can scale errors rapidly. This article is educational and not investment or financial advice.

Frequently asked questions

Is classification or regression better for trading?

Neither is universally better. Classification fits discrete outcomes or threshold events; regression fits magnitude or fair-value estimates. The correct target is the one that maps cleanly to an executable decision after costs and risk.

Why is direction accuracy not enough?

Direction labels ignore move size. A correct “up” prediction can lose money when the gain is smaller than spread, fees, slippage, funding or market impact.

What should a trading model predict?

Useful outputs include a return distribution, executable fair value, probability that a cost-adjusted threshold will be crossed, or expected utility under explicit position rules. The horizon and reference price must be defined.

What is probability calibration?

A model is calibrated when events assigned roughly 70% probability occur about 70% of the time across comparable forecasts. Calibration should be checked out of sample and over time, not inferred from confident wording.

What is an abstain or no-trade state?

It is an explicit decision not to trade when estimated edge is smaller than costs, uncertainty or a risk threshold. Forcing a binary buy/sell choice can turn weak forecasts into guaranteed turnover.

Should transaction costs be part of the target?

They must at least be part of evaluation and decision logic. If costs vary with size and market state, a gross-return target can be modeled separately, but the trade rule must use realistic executable costs.

Why are overlapping labels dangerous?

They make neighboring samples share much of the same future price path. Random train/test splits can then leak nearly identical market episodes across both sets and overstate performance.

Which metrics are better than accuracy?

Depending on the task, use log loss or Brier score for probabilities, MAE or distributional scores for returns, calibration plots, and a fully specified net-P&L simulation with drawdown, turnover and capacity checks.

Does a high backtest Sharpe ratio prove the model works?

No. Repeated testing, selection bias, leakage, unrealistic fills and a short favorable regime can inflate results. Preserve a final untouched time period and progress through replay and paper trading before risking capital.

Can an LLM decide position size from its confidence?

Not safely from prose confidence alone. Position sizing should come from deterministic rules using calibrated forecasts, current exposure, liquidity, loss limits and verified market data.

Should a Trading Model Predict Up or Predict a Price?