Skip to content
Crypto.world

Research

A novel approach to trading strategy parameter optimization using double out-of-sample data and walk-forward techniques

Most trading-strategy papers optimize indicator parameters on the full sample and then repeatedly "test" on data they have already effectively seen, which inflates reported performance and hides…

A novel approach to trading strategy parameter optimization using double out-of-sample data and walk-forward techniques
arXiv
2602.10785
Published
2026-02-11
Authors
Tomasz Mroziewicz, Robert Ślepaczuk

AI summary

Overview

Research area: Quantitative finance / algorithmic trading (q-fin.TR), specifically the methodology of backtesting and parameter optimization for intraday cryptocurrency strategies.

Technical level: Advanced. The paper assumes familiarity with walk-forward optimization, Sharpe-type risk metrics, moving-average signals and block bootstrap resampling.

Scope in one sentence: The paper treats the lengths of walk-forward training and testing windows as a tunable hyperparameter, selects window combinations on a 19-month Bitcoin training period, and then evaluates the two best combinations exactly once on a 21-month unseen period for Bitcoin, Binance Coin and Ethereum.

What This Paper Is About

Most trading-strategy papers optimize indicator parameters on the full sample and then repeatedly "test" on data they have already effectively seen, which inflates reported performance and hides overfitting. The authors ask whether a strategy's results depend on the arbitrary choice of walk-forward window lengths, and they build an evaluation pipeline in which the final holdout is touched only once. The goal is a practical framework that combines profitability estimation, statistical significance testing and overfitting control, demonstrated on intraday Bitcoin data with the Exponential Moving Average (EMA) Crossover strategy.

Key Contributions

  1. Window length becomes an optimization target, not an assumption. Instead of fixing walk-forward window sizes arbitrarily, the authors run 81 combinations of training and testing window lengths spanning 1 to 28 days (1, 2, 3, 5, 7, 10, 14, 21 and 28 days, with multiples of 7 included to probe weekly seasonality), each producing its own Sharpe ratio from walk-forward testing returns. Results are arranged as a grid (training length on one axis, testing length on the other) and smoothed with a weighted rule from Gómez and Ślepaczuk (2021), where the original cell keeps half the weight and the rest is distributed over neighbors with weights 1/n.

  2. A "double" out-of-sample architecture. Every reported result comes from data not used for fitting: first the aggregated walk-forward testing segments over the Global Training Data Period (February 8, 2018 to September 1, 2019), then a single evaluation on the Unseen Data Period (November 7, 2019 to August 22, 2021), used only once, with no retraining.

  3. Cross-asset transfer of parameters. The two window combinations with the highest smoothed Sharpe ratios were estimated on Bitcoin and then applied, unchanged, to Binance Coin and Ethereum to test adaptability.

  4. A full evaluation package around the strategy. This includes a portfolio that blends Buy-and-Hold with the walk-forward strategies, a cost sensitivity analysis over seven fee levels (0.05%, 0.07%, 0.10%, 0.20%, 0.30%, 0.40%, 0.50%), and two block-bootstrap procedures (random EMA-pair selection and shuffled position blocks, 1,000 iterations each) applied to the top two window combinations.

Main Findings

  • Window size materially changes results. Strategy performance under walk-forward optimization based on the Robust Sharpe Ratio is highly dependent on the chosen training and testing window lengths, which the authors frame as a tension with the Efficient Market Hypothesis (Fama, 1965), under which such variation should not matter.

  • All 81 window combinations beat Buy-and-Hold in training. During the Bitcoin Global Training Data Period, every combination of walk-forward windows outperformed a Buy-and-Hold benchmark.

  • Out-of-sample parity with lower risk. In the Unseen Data Period, the strategy performed similarly to Buy-and-Hold but with lower drawdown and a higher Information Ratio.

  • The result transfers to other coins. Similar outcomes were observed for Binance Coin and Ethereum when the Bitcoin-derived window parameters were applied.

  • The portfolio is the strongest result. A portfolio combining Buy-and-Hold with the walk-forward strategies outperformed all individual strategies and Buy-and-Hold alone, achieving the highest overall performance and a 50% reduction in drawdown.

  • Persistence without retraining. The strategy delivered positive returns for nearly two years on intraday cryptocurrency data without being retrained.

  • Costs are survivable but binding. A conservative fee of 0.1% per transaction was used throughout; a long-to-short turnaround costs 0.2% (entry plus exit) at that level and 1% at the highest tested fee of 0.5%. The cost sensitivity analysis put the break-even point at roughly 0.4% per transaction.

  • Data are strongly non-normal. Bitcoin returns across the six sampling intervals (1, 5, 10, 15, 30 and 60 minutes) show high kurtosis with Jarque–Bera p-values of 0. Kurtosis is 18.13 at 1 minute and declines with the sampling interval, except at 5 minutes, where it rises to 31.97; the authors note prior work (Petukhina Alla A. and Karl, 2021) documenting Bitcoin 5-minute kurtosis as high as 49. The paper's text states that all intervals show positive skew, while Table 1 reports negative skew values at several intervals (−0.0769 at 1 minute, −0.0324 at 10 minutes, −0.3408 at 15 minutes).

  • Statistical testing setup. The null hypothesis is that the strategy has no statistically significant advantage over randomly generated trading signals; at a 5% confidence level the strategy must beat 95% of random strategies. The numeric bootstrap results are not included in the portion of the paper available here.

Methodology in Plain English

The researchers take minute-level price data from a Kaggle "G-Forecast" challenge dataset covering 10 cryptocurrencies, and work with a subset at six sampling frequencies from 1 to 60 minutes. The timeline is split into a 19-month Global Training Data Period and a 21-month Unseen Data Period, with a gap between them so that both start and end at comparable calendar points after the walk-forward adjustment.

The trading rule is simple: compute a fast and a slow EMA (from the set 5, 7, 10, 15, 20, 30, 40, 50, 100, 150 and 200 periods, with anything above 35 treated as "slow"), go long when the fast EMA is at or above the slow EMA and short otherwise.

For each of the 81 training/testing window pairs, the engine slices the training period into consecutive blocks. Inside each block it tries every fast/slow EMA pair on the training slice, keeps the pair with the highest Sharpe ratio, and then applies that pair to the immediately following test slice. Only the test-slice returns are stitched together to compute that window combination's overall Sharpe ratio, so the in-sample data never contributes to the reported performance. The resulting grid of Sharpe ratios is smoothed because neighboring cells vary erratically, and the two highest cells are selected.

Those two window combinations are then run once, without modification, on the unseen period for Bitcoin, Binance Coin and Ethereum. To check whether the edge could be luck, the authors resample with a block bootstrap suited to autocorrelated, heteroscedastic time series: one variant compares the optimized EMA choice against randomly chosen EMA pairs, and another shuffles blocks of consecutive long/short positions and re-matches them to actual returns. Cost sensitivity is handled by re-pricing the frozen list of positions at each of seven fee levels, holding everything else constant.

Why This Matters

Impact on research. The paper directly attacks two practices the literature review flags as endemic: repeated optimization on data labeled out-of-sample, and neglect of transaction costs (the review notes that only 37% of studies surveyed by Farias Nazário et al. (2017) adjust for risk, and Park and Irwin (2007) list data snooping, rule selection and cost estimation among persistent biases). Treating window length as a documented hyperparameter, and reserving a genuinely single-use holdout, gives other researchers a reproducible template rather than an ad hoc setup.

Real-world applications.

  • Auditing or designing backtests for crypto and high-frequency strategies where the choice of rolling window is usually undocumented.
  • Setting realistic fee and slippage assumptions: the finding that the strategy breaks even near 0.4% per transaction tells a practitioner at what venue cost level the whole approach stops working.
  • Constructing blended portfolios that pair passive Buy-and-Hold exposure with an active EMA-walk-forward overlay, which is where the largest reported drawdown reduction appears.
  • Transferring parameters across correlated assets (Bitcoin-derived settings applied to Binance Coin and Ethereum) instead of re-optimizing each market separately.

Industry relevance. The framework speaks to anyone shipping systematic strategies that must survive live markets: it quantifies how much of a backtest's appeal may come from window choice, shows that a mechanically simple rule can reduce drawdown relative to Buy-and-Hold, and demonstrates a cost threshold beyond which intraday execution economics fail.

Future Directions

  • Better selection rules for window pairs. The authors themselves propose choosing the combination that minimizes the gap between training and testing performance, or using bootstrapping to pick the combination with the highest statistical significance, instead of simply taking the top smoothed Sharpe cells.
  • Periodic retraining during the test period. Because the strategy stayed profitable for nearly two years without retraining, the authors raise retraining as an obvious next experiment, alongside the risk of parameter decay over time.
  • Sensitivity to the train/unseen split. The paper acknowledges that a different division between the Global Training Data Period and the Unseen Data Period could yield different optimal parameters; how stable the chosen windows are across alternative splits remains open.
  • Accounting for data mining at the window level. Searching many training/testing combinations while also selecting the best EMA inside each training block is flagged as a residual overfitting risk, inviting methods that correct explicitly for this multi-level search.

Target Audience

Quantitative researchers and graduate students in finance who build or review backtesting pipelines; practitioners at systematic funds and crypto trading firms who need defensible validation before capital is committed; and journal referees or risk officers evaluating whether a reported strategy's performance is genuine or an artifact of window choice and repeated holdout use. Readers without a background in time-series resampling and risk-adjusted performance metrics will find the methodology section demanding.

Authors’ abstract

This study introduces a novel approach to walk-forward optimization by parameterizing the lengths of training and testing windows. We demonstrate that the performance of a trading strategy using the Exponential Moving Average (EMA) evaluated within a walk-forward procedure based on the Robust Sharpe Ratio is highly dependent on the chosen window size. We investigated the strategy on intraday Bitcoin data at six frequencies (1 minute to 60 minutes) using 81 combinations of…

The full abstract on arXiv
Read the original paper