A walk-forward crypto system, and the seven ideas it killed

Paper-traded research for methodological transparency. Not investment advice.

Why this exists

I'm a backend/AI engineer. I wanted to know what happens when you take the boring, honest version of quantitative trading — walk-forward testing, real costs, real funding, negative results published — and apply it to crypto with only free data. This project is the answer, including the parts that failed. It doubles as my working sample for trading-infrastructure roles and as the research base for a crypto-data product I'm building.

How it works

Two coins, BTC and ETH, each running an independent book, equal capital.

Every Monday at 00:00 UTC, each book re-ranks a fixed zoo of 24 simple configurations — four families: trend (moving-average crosses, in long/short, long-only and short-only variants), breakout (channel breakouts), mean-reversion (oscillator and band signals), and cash — by their Sharpe ratio over the trailing 30 days of 4-hour bars, net of costs and funding. It then trades the equal-weight average of the top three for the next week, out of sample. If nothing scores positive, the book sits in cash.

Every Monday at 00:00 UTC:

1. Score all 24 configurations on the trailing 30 days of 4-hour bars, net of costs and funding.
2. If nothing scores positive, hold cash for the week.
3. Otherwise trade the equal-weight average of the top three for one week, out of sample.
4. Scale the position to a 50% annualized volatility target, capped at 1.5x.
5. Repeat next Monday.

Two structural choices do most of the risk work:

• Top-3 averaging instead of top-1. The single best trailing config is the most overfit pick. I learned this the embarrassing way: the top-1 version's yearly result moved by 12 percentage points when I changed how exact score ties break. The top-3 basket is immune to that by construction.
• Volatility targeting. Positions scale to a 50% annualized volatility target with a 1.5× cap, so size grows in calm trends and melts away when the market turns violent. The same machine without the cap headroom returns +30.2%/+37.4%/+12.6% across 2026/2025/2024 — the dial buys ~1.4× the size at the same Sharpe, and proportionally deeper drawdowns.

Costs: 0.15% per side on every position change, charged on the net position delta. Funding: real Binance funding events, timestamped and signed (shorts receive when longs pay). No price data is assumed beyond public Binance history, cross-checked against Bybit and OKX.

Results

Per-year frozen replays (selection restarted each Jan 1, settings frozen):

What these numbers don't prove

• 2026 is the development period. Design decisions were made while watching 2026 unfold. The machine's own settings (bar size, refit cadence, training window) were chosen by testing 108 combinations on this same year — only 25% of those cells made money in 2026 and the median cell lost ~18%. The clean evidence is the frozen 2024 and 2025 replays, which the machine never saw during design.
• The 2026 Sharpe of 2.25 has a standard error of ±1.5 (5.4 months of data). The 95% interval spans roughly −0.7 to +5.2. The full-period Sharpe of 1.20 ± 0.64 is the number I'd plan around, and the true forward expectation is plausibly lower still.
• 2024 underperformed buy-and-hold badly (+16.5% vs BTC's +121%). This system earns its keep in flat and falling markets; in a straight bull you'd rather just hold.
• Refit-day luck is real. Re-running with the weekly refit on a different weekday moves yearly results by up to ~33 points. Monday ranks top-2 in all three years and is the natural week boundary, but single-year numbers deserve that grain of salt.
• These are paper-trading research results, not audited returns.

The graveyard

Every idea below sounded reasonable. Each was implemented, tested across 2024, 2025 and 2026, and rejected. Deltas are recomputed fresh against the shipped baseline (sources in the numbers ledger).

1. Refit daily instead of weekly. Why it sounded good: markets move fast; re-fit faster, adapt faster. What the data showed: −33.7, −40.9 and −25.9 points across the three years. A 30-day ranking barely changes in a day; daily refits just churn positions and pay costs. Adaptation speed and decision quality are different things.

2. Daily bars instead of 4h. Why it sounded good: fewer bars, less noise, cheaper. What the data showed: −11.8 / −45.7 / −14.8 points. With daily bars a 30-day window is only 30 observations — too few to rank 24 configs — and exits lag by days. (1-hour bars, for the record, were not clearly worse at current settings — mixed deltas of −8.6/+6.5/−1.4 — just more churn for no consistent gain. I keep 4h.)

3. A drawdown kill switch. Why it sounded good: cap the pain — go flat after an X% drawdown. What the data showed: the worst idea tested. On the current baseline it costs up to −43 points (2026); on an earlier variant it flipped 2025 from +41% to −12.6%. Crypto drawdowns end in V-shaped reversals; a kill switch systematically sells the bottom and watches the recovery from cash.

4. Drawdown-triggered emergency refits. Why it sounded good: don't go flat, just re-decide immediately when losing — with a confirmation rule and a cooldown. What the data showed: triggers tight enough to fire still fire at local bottoms (−22 points in 2026); triggers loose enough to be safe never fire, because vol targeting already contains the episodes. The weekly schedule plus vol scaling is the adaptation mechanism.

5. Sentiment and positioning signals in the zoo. Why it sounded good: more information, better picks — Fear & Greed and funding-positioning signals as extra candidates. What the data showed: −13.6 and −22.8 points in 2024/2025 (2026 alone improved slightly — which is exactly how overfitting bait looks). Every extra candidate makes a 30-day ranking easier to fool: the selection step pays for information in noise.

6. Fixed leverage. Why it sounded good: the edge is positive, so multiply it. What the data showed (in a separate study with intraday liquidation simulation): every fixed-leverage configuration at 1.5× and above eventually hit a day that zeroed the account — most of them on 2020-03-12 — after triple-digit best years. Leverage without volatility scaling isn't a multiplier, it's a countdown. The shipped 1.5× cap works because vol targeting shrinks exposure before storms, not after.

7. "Know the winners, size them up." Why it sounded good: if you can tell good trades from bad ones in advance, lever the good ones. What the data showed: across 126 measured weeks the hit rate is 52% — the system earns through asymmetry (average winning week +4.1%, average losing week −2.7%), not foresight. Six confluence overlays — basket unanimity, conviction score, cross-book agreement, trend-efficiency, calm-trend gating, and real positioning data (open interest, top-trader long/short, taker flow) — all failed to beat a flat cap. Nothing I could measure predicts which trades win.

Engineering notes

• Free deep-history data is available if you know where it lives. Binance's REST stats endpoints cap at ~30 days, but the data.binance.vision bulk archive serves the same positioning metrics (OI, top-trader and global long/short, taker ratio) as daily files — I backfilled 923 days at 4h resolution with zero gaps. The rest of the free stack: Deribit DVOL (implied vol, 2021→), DefiLlama stablecoin supply (2020→), CFTC COT for CME bitcoin futures (weekly, ~2018→), Coinalyze daily aggregates, and basis reconstruction from Binance continuous/premium-index klines. None needs a paid key.
• Clean your klines. Binance history contains real artifacts: I found a $0.0001 print (an empty-order-book wick on Black Thursday 2020) that falsely "liquidated" even unlevered backtests until liquidation marks were floored to index-style prices. The 4h grid also has eight maintenance gaps in 2018–2020 (none after 2023). Cross-checking against a second venue is cheap insurance — Binance vs OKX closes agree to a median 0.6 basis points, which also proves the strategy isn't a data-feed artifact.
• Funding must be interval-aware. Binance changed its funding formula on 18 Sep 2025 (variable 1/2/4/8h intervals). This pipeline applies individual timestamped funding events to the bar containing them, so it's interval-agnostic by construction; BTC and ETH stayed on 8h schedules through the change (verified by event counts).
• 0.15%/side is deliberately conservative. Binance VIP-0 is 0.1% spot / 0.05% perp-taker before BNB discounts, and slippage on BTC/ETH at retail size is a rounding error. Real execution should beat the backtest's cost assumption, not miss it.

What I'd test next — status report

• Top-k ensembling — tested and adopted. Averaging the top-3 ranked configs replaced top-1 after the tie-break fragility discovery; it turned the untuned years from +11%/−2% to +37%/+13% (unleveraged basis) while giving up a little of the development year.
• A noise-aware selection hurdle — tested, rejected. The max of 24 correlated trailing Sharpes is positively biased, so I swept hurdles (winner must clear τ); τ=2 produced the best 2026 on record and gutted 2025. The plateau (top-3 averaging) handles selection noise better than a gate.
• Switching hysteresis — tested, rejected. Making the weekly switch lazier (incumbent keeps its seat unless beaten by a margin) was a wash at small margins and worse at large ones.
• Rebalance-day tranching — measured, not adopted. The weekday sweep shows large dispersion (~33pp), which argues for splitting capital across refit days; but Monday is consistently top-2 and tranches would trade away expected return for path smoothness I don't currently need. Worth revisiting with real capital.
• Range-based vol estimators — partially open. EWMA vol was tested and rejected (the 30-day estimator's lag usefully keeps size small through post-spike chop). Parkinson / Yang-Zhang range estimators on 4h OHLC remain genuinely untested.

FAQ

Does walk-forward optimization work for crypto?
In this experiment: yes, with weekly cadence, a 30-day window, and brutal honesty about costs. But 75% of the 108 design combinations I tried lost money in 2026. Walk-forward is a method, not an edge — most settings of it are still wrong.

How do you avoid overfitting with 24 strategies in the pool?
Three ways: trade the average of the top three instead of the single winner; judge every change on years it wasn't designed on (2024/2025 frozen replays); and publish the rejections. The pool itself never grew — adding candidates was tested and made things worse.

Is 0.15% per side realistic on Binance?
It's conservative. VIP-0 taker fees are 0.1% spot / 0.05% USDT-perp before discounts, and BTC/ETH books absorb retail size with negligible slippage. The assumption deliberately overcharges so live results should beat the backtest.

Why only BTC and ETH?
Tested. Adding the largest alts (SOL, BNB, XRP, DOGE — equal-weighted books) collapsed the development year from +45% to roughly flat in the worst variant: weak books steal capital from strong ones. The edge concentrates in the most liquid, cleanest-trending pairs.

Why didn't a stop-loss / kill switch help?
Because crypto drawdowns resolve as V-shaped reversals more often than as slow bleeds. Every drawdown-triggered rule I tested — kill switch, emergency refit, confirmation-gated variants — sold bottoms. Sizing by volatility before the storm beat reacting after it, every year.

Can I see which signals it's using right now?
Family-level, yes — the live dashboard and the weekly ribbon chart show trend/breakout/mean-reversion/cash per coin per week. Exact parameters are deliberately not published (see the redaction report in the research repo).

Is this live money?
No — paper trading against live Binance prices, with fsync'd append-only trade logs and a public dashboard. The point is the method and the record, not a brokerage statement.

Work with me

I build trading infrastructure, data pipelines and research tooling like this for a living. Available for trading-infrastructure and data engineering work. Drop me a line or find me on LinkedIn.

You can also view the Source and methodology on GitHub or explore the Interactive case study.

Everything here is educational and informational only. These are paper-trading research results, not investment advice and not an offer to manage money. Past performance — simulated or otherwise — does not guarantee future results. Crypto trading can lose more than you invest.