How much historical data should you train on?

Every walk-forward pipeline has to answer one question before the first fold ever runs: how much history goes into training. Too little and the model chases noise. Too much and it drags in a market regime that no longer applies. Most teams settle this by habit rather than evidence, usually whatever window fit in a first experiment, and never revisit it.

We wanted a direct answer, tested rather than assumed. Does a longer lookback actually produce a better model, or does it just feel more rigorous? Does the answer depend on the market, the model family, or how often you retrain? And separately, whichever training length wins on prediction, does it also win on the P&L line, once real spreads, fees, slippage and funding are subtracted?

This article lays out how we tested it, across two very different asset classes, three model families and six training lengths, and what we found.

We trained the same models on six different lookback windows, from one calendar month up to two years, and evaluated each one on the bar directly after the training window ended. To make sure the comparison wasn't an artifact of one particular evaluation choice, every training length was scored under three separate out-of-sample schedules: a fixed one-month test, a fixed three-month test, and a test window sized to two-thirds of the training length, a rule of thumb some practitioners use to keep the train and test ratio roughly constant as the window grows.

The universe spanned two markets that behave very differently: a basket of large-cap US equities and a basket of liquid Binance perpetual contracts, each sampled at two bar frequencies. Three model families were trained side by side: a regularized linear model, a gradient-boosted tree ensemble and a small neural network, all fed the same technical feature set built purely from price and volume. No model ever saw a cost, fee or funding figure as an input; those only entered later, at the P&L stage.

Every training length, market, frequency and model combination was walked forward across roughly three and a half years of out-of-sample history, refitting on schedule the whole way through, so this is never a single snapshot. Selection inside each window used a fixed train and validation split and a locked hyperparameter grid, so no training length ever got a search-budget advantage over another. The full run stacked up to 5,868 rolling model-windows and 176,040 total configuration attempts.

We then translated the winning configuration in each window into an actual position: long the strongest predictions, short the weakest, sized and executed exactly as a systematic desk would, with measured spreads, real exchange fees, slippage and signed funding subtracted from every fill. That step separates a model that predicts well from a model that would have made money, which are not the same claim.

Longer training history produced better predictions almost everywhere. Under the fixed one-month schedule, pooled rank IC rose from 0.0068 at one month of training to 0.0173 at 24 months, with the gain concentrated below three months and a shallow plateau between two and six months. The three-month schedule shows the same pattern, rising from 0.0063 to 0.0143.

Twenty-four months won every fixed-window comparison under the one-month and three-month schedules, and tied three months on cross-cell wins under the two-thirds schedule, where six months also came close to the 24-month median. Tallied across all 36 market, timeframe, model and schedule cells on pooled predictive IC, 24 months is the outright winner, with three months a distant second and every shorter length splitting the rest.

The aggregate hides some texture. Shorter windows still beat 24 months in a handful of individual cells, between two and four cells depending on the schedule, after correcting for multiple comparisons with Benjamini-Hochberg. Twenty-four months was significantly worse in none. The aggregate result holds because the long window's wins were larger and more frequent than its occasional losses, not because it won everywhere.

Crypto and equities gave different answers. Crypto was the clean case: 24 months won every single pooled-IC comparison across both frequencies and all three model families under the fixed schedules, with winning ICs as high as 0.0526. Equities never converged on one answer. Three months won half of the fixed-window equity cells, and the strongest equity IC reached only 0.0129, a fraction of the best crypto result. If you only train one model on one market, this is the result worth internalizing: “use more history” is close to a law in crypto, and a much softer default in equities.

None of it survived transaction costs. We converted every window's predictions into an actual long-short book, entering at the next bar's open and exiting at that bar's close, and charged it real costs.

Crypto costs ran to 3.7 bps per selected position against 0.05 bps of gross P&L; equity costs ran to 0.8 bps against 0.03 bps gross. And the training length that predicted best did not rescue the economics, because the cost gap was never close enough for forecast quality to matter.

Not one of the 180 shorter-history-versus-24-month net P&L contrasts we ran was statistically distinguishable from zero after correcting for multiple testing, in either direction. Training history changes forecast quality. It does not turn this particular next-bar ranking strategy into a profitable one.

Every leak, timing and integrity check we ran against this result passed; the exact checks and their numbers are on the reproducibility page.

• The result describes this fixed basket of 20 instruments over this roughly 3.5-year window. It is not a claim about either asset class as a whole.
• Crypto trades around the clock; equities have overnight gaps. The two markets share a training and testing calendar but not the same bar economics.
• One instrument in the crypto basket stops trading partway through the sample and was kept in place rather than removed, to avoid survivorship bias, which makes that panel deliberately ragged.
• The economic result is specific to a next-bar long-short ranking rule at fixed position weights. A different holding period, execution style or portfolio construction is a different test.
• Twenty-four months is the strongest default given the choices tested here, not a mathematical optimum. Shorter windows still win some individual cells.