Bitcoin Walk-Forward Validation Lab

How the Bitcoin Walk-Forward Validation Lab works

Split Bitcoin’s weekly history into consecutive test folds, fitting the power-law model in each only on earlier data. The Balanced strategy runs fresh in every fold against same-cash-flow DCA, showing whether it only worked in one market episode.

What you enter

• Minimum training years (3–10; default 5)
• Test months per fold (3–36; default 12) and step months (3–36; default 12)
• Maximum folds (2–24; default 10; most recent kept)
• Fixed strategy: buy at percentile ≤ 25 (50% of cash), sell at ≥ 90 (25% of BTC), 10,000 USD capital, 500 USD per month, 10 bps fee, 5 bps slippage

What it shows

• Valid test folds and folds beating DCA with their share
• Median excess with P25–P75 range, and worst-fold excess
• Per-fold table: training end, test period, strategy, DCA, excess, drawdown and trades
• Fold CSV (adds return on contributions) and a validation snapshot

How to use it

1. Choose enough training years for the first fold.
2. Set fold length and step.
3. Cap the fold count; recent folds are kept.
4. Review win share, median and worst fold, then export or snapshot.

Formula

The first fold starts at the first observation + minimum training years; later folds advance by the step (1 month = 30.4375 days) and last the test months. Each fold runs the power-law backtest over its window, refitting monthly on all earlier data. Excess = strategy final value − DCA final value. Folds with fewer than 8 weekly observations are dropped; at least 2 valid folds are required.

Worked example

Six folds with excess (USD) +1,200, −300, +450, +2,100, −800 and +150: 4/6 ≈ 66.7% beat DCA. Sorted, the median = (150 + 450) / 2 = 300 and the worst fold = −800.

How to read the result

• Many winning folds and a mild worst fold suggest the edge is not one episode.
• If nearly all gains come from one fold, the full-history result can mislead.
• A step shorter than the fold length makes folds overlap.

Limits and cautions

• Out-of-sample testing reduces but does not remove overfitting or regime dependence.
• Each fold restarts at 10,000 USD, so values do not carry over.
• Parameters are fixed at Balanced; backtester changes do not apply here.

Common questions

What is walk-forward validation?

Testing on successive slices of history where each slice’s model uses only earlier data.

Why does it say history is insufficient?

It needs 160 weekly observations and 2 valid folds. Lower training years, fold length or step.

Does the training window slide?

No, it expands from the first observation to each fold’s start.