A training set is the collection of observations used to fit a model. In supervised learning, each observation pairs features with a target, such as the return following a trading decision. For a quant strategy, a row belongs in a historical fit only if its required information was available when that fit would have occurred. An old row date alone does not establish that its outcome was known.
Training determines fitted quantities: regression coefficients, tree splits or neural-network weights, for example. A validation set helps choose settings or competing models; a test set assesses the chosen procedure on held-out observations. A low training error describes how well the model fits the data it learned from. It does not establish useful future prediction or a profitable trading strategy.
Sixty allocated rows can mean fifty-seven usable rows
Consider 100 synthetic decision rows, ordered in time. Allocate the first 60 to candidate training observations, the next 20 to validation and the final 20 to testing. These proportions are an illustration, not a generally recommended 60/20/20 rule. Useful window sizes depend on the prediction horizon, data dependence, model complexity and the amount of history available.
Now suppose row has a three-step forward-return label, which becomes usable immediately at time . Fit the model just after all time-60 publications have arrived, before the decision at time 61. Assume every feature was constructed from information available at its own decision time.
The training-label condition is
Only rows 1–57 qualify. Row 57's label arrives at time 60 and is usable under the stated timing convention. Rows 58–60 have label-availability times 61, 62 and 63; exclude them from this fit. Thus there are 57 eligible training observations, even though the nominal training allocation contained 60 rows. Their outcomes may later become available for a later fit under a predeclared updating schedule.
Original synthetic timeline. Labels become usable immediately at time i + 3; fitting occurs after time-60 outcomes arrive. The 60/20/20 allocation illustrates roles rather than recommending a universal split.
This boundary check prevents a form of look-ahead bias. Simply shuffling financial rows before splitting can also mix later information into earlier fits. A chronological allocation still needs checks for outcome overlap, publication delays and the actual forecasting task.
The scaler learns too
For this example, fit the scaler, imputation rules and any data-driven feature selection on the 57 eligible training rows. Use those fitted transformations on validation and test inputs. Computing a scaling mean from all 100 rows would allow the later data distribution to influence the fitted pipeline, even if the prediction model itself sees only earlier rows.
Each fitting window in cross-validation needs its own fitted preprocessing. Separately, choosing which preprocessing method or feature subset performs best is a model-selection decision: it belongs inside the validation design. Repeatedly changing the pipeline after inspecting test results makes that test part of development and invites overfitting.
The Time-Ordered Train, Validation and Test Design lesson develops this learning schedule, including eligible labels and rolling or expanding fits. Its public preview introduces the lesson; full access requires membership.
More rows are not necessarily more independent evidence
Here, “57” is a count of eligible rows. It is not a claim of 57 independent observations. Overlapping return labels share price changes; assets observed on the same date can share market shocks; repeated observations can cluster within regimes. These forms of dependence affect how much information the training data contain and how uncertain an assessment is.
There is no universal conversion from row count to independent sample size. More intraday rows may add less new information than a longer span covering different conditions. Conversely, a longer training history can introduce stale relationships when the market changes. Choosing a training window therefore trades off estimation stability, relevance and the risk of fitting noise; counting rows is only the beginning.
Sources
- Hastie, Tibshirani and Friedman, The Elements of Statistical Learning, second edition, §7.2, p.222: distinguishes fitting, selection and assessment, and explains why split sizes depend on the problem. Its general random-split illustration should not be treated as a finance chronology rule.
- scikit-learn, Common pitfalls and recommended practices, §§12.1–12.2: learn preprocessing on training data, then apply the same fitted transformation to held-out data.
- scikit-learn, Cross-validation of time series data: explains why temporally correlated observations require a different splitting approach from ordinary random partitions.