QUANT / LAB · CRYPTO RESEARCH
Can crypto CNNs trade less?
We explored daily prices for eight cryptocurrencies. Daily CNN rebalancing incurred substantial assumed costs; trading less helped in some windows, but a stable advantage remains unproven.
This is retrospective research, not a fresh holdout, live performance or competition returns. The fixed 2023–2025 universe has selection bias; fills and asset events remain unverified. The data and results have not passed formal research acceptance.
Explore a window’s return paths
Choose a window, cost and method to compare it with equal-weight buy and hold. Each starts at 1 USDT and ends after liquidation costs. The vertical scale changes with the selection.
The middle outcome across eight windows
These medians assume a 0.35% one-way total cost per trade, not a verified exchange fee. Each window starts in cash and closes its positions. They are not the return of one continuous three-year account.
| Method | Window return | Maximum drawdown | Cumulative turnover multiple |
|---|---|---|---|
| Equal-weight buy and hold | -3.68% | 25.49% | 2.00× |
| Equal-weight daily | -3.47% | 25.65% | 3.23× |
| Equal-weight with volatility control | 1.89% | 17.29% | 3.89× |
| Momentum, top three | 3.58% | 29.06% | 12.95× |
| Momentum with volatility control | 2.08% | 17.10% | 10.80× |
| Numeric CNN, daily | -9.17% | 32.74% | 45.32× |
| Numeric CNN, weekly | -4.03% | 30.91% | 13.78× |
| Numeric CNN, every five days | -1.60% | 27.95% | 16.15× |
| Numeric CNN, rank buffer | -7.96% | 27.09% | 19.33× |
| Numeric CNN with volatility control | -5.75% | 20.78% | 32.80× |
| Candlestick-image CNN | -7.06% | 29.96% | 32.18× |
Three later windows in 2026
We kept the earlier rules and compared equal weight, momentum and CNN. Each window starts in cash, liquidates at the end and assumes a 0.35% one-way total cost.
These dates had previously been used in other research, so they are not a pristine holdout. Data has not passed formal research acceptance, and the next daily open can be optimistic.
In all three windows, equal weight and momentum with the same target-exposure control had higher returns than the controlled CNN. Three windows do not establish what will happen next.
See each window’s return, drawdown and turnover
| Test window | Method | Window return | Maximum drawdown | Cumulative turnover multiple |
|---|---|---|---|---|
| 2026-01-06 — 2026-03-31 | Equal-weight buy and hold | -37.40% | 41.36% | 2.00× |
| 2026-01-06 — 2026-03-31 | Equal-weight with volatility control | -30.70% | 32.38% | 3.95× |
| 2026-01-06 — 2026-03-31 | Momentum, top three | -47.29% | 48.89% | 19.21× |
| 2026-01-06 — 2026-03-31 | Momentum with volatility control | -36.14% | 37.74% | 13.47× |
| 2026-01-06 — 2026-03-31 | Numeric CNN, daily | -47.62% | 49.42% | 42.47× |
| 2026-01-06 — 2026-03-31 | Numeric CNN with volatility control | -38.50% | 40.23% | 28.71× |
| 2026-01-06 — 2026-03-31 | Numeric CNN, weekly | -36.09% | 41.49% | 12.75× |
| 2026-04-06 — 2026-06-29 | Equal-weight buy and hold | -20.35% | 31.62% | 2.00× |
| 2026-04-06 — 2026-06-29 | Equal-weight with volatility control | -19.92% | 30.86% | 3.89× |
| 2026-04-06 — 2026-06-29 | Momentum, top three | -24.50% | 33.63% | 13.85× |
| 2026-04-06 — 2026-06-29 | Momentum with volatility control | -22.47% | 31.30% | 12.52× |
| 2026-04-06 — 2026-06-29 | Numeric CNN, daily | -38.24% | 43.28% | 60.32× |
| 2026-04-06 — 2026-06-29 | Numeric CNN with volatility control | -35.12% | 40.54% | 54.38× |
| 2026-04-06 — 2026-06-29 | Numeric CNN, weekly | -29.21% | 39.14% | 14.15× |
| 2026-07-05 — 2026-09-25 | Equal-weight buy and hold | 40.18% | 9.80% | 2.00× |
| 2026-07-05 — 2026-09-25 | Equal-weight with volatility control | 33.78% | 6.45% | 3.37× |
| 2026-07-05 — 2026-09-25 | Momentum, top three | 26.52% | 11.90% | 17.21× |
| 2026-07-05 — 2026-09-25 | Momentum with volatility control | 23.22% | 11.20% | 14.72× |
| 2026-07-05 — 2026-09-25 | Numeric CNN, daily | 18.30% | 17.06% | 45.78× |
| 2026-07-05 — 2026-09-25 | Numeric CNN with volatility control | 12.52% | 16.74% | 41.47× |
| 2026-07-05 — 2026-09-25 | Numeric CNN, weekly | 26.11% | 10.19% | 12.75× |
Cumulative turnover adds one-way notional divided by account value at each trade; it is not a trade count. The same volatility target sets target exposure, not necessarily the same realized risk.
A sentiment snapshot exists, but its timing remains uncertain
A BTC and ETH sentiment history was retrieved on 2026-10-14. An exploratory filter covered only those two assets; its incremental effect against an equal-target-exposure comparison was inconsistent.
First publication times and historical revisions are unverified. The assumed 24, 48 and 72 hour waits do not prove what was available at the time; sentiment was not invented for other assets.
Why historical asset identity needs checking
The MATIC to POL migration had a trading gap; old LUNA became LUNC and a new asset reused LUNA. A date or ticker alone does not establish identical asset identity or trading hours. These examples are not a complete historical universe.
MATIC and POL announcement ↗ · Old and new LUNA announcement ↗
A future check has been specified for November 2026 through January 2027. The data has not yet been generated or acquired, and there is no result or automatic tracking task.
How to read the results
- The volatility overlay matches target exposure; realized risk can still differ across portfolios.
- The image model predicts absolute returns; the numeric model ranks coins. Differences cannot be attributed to image input alone.
- Signals use the daily close and simulated fills use the next daily open. Crypto trades continuously, so the adjacent open can be optimistic; an extra day's delay also changes the holding period.
- The fixed eight-asset universe is not a point-in-time market history. Token events, source data and actual execution conditions remain to be checked.
- The 2023–2025 comparison used no sentiment data; historical availability and current access to the later snapshot remain unverified.
- These are feasibility settings. Each image model trains on just 200 images for two epochs. Results need new data for further testing.
What we did
We used 2023–2025 UTC daily spot data for eight USDT pairs from Binance monthly archives. Source and historical availability have not been independently accepted. The numeric CNN uses a 400-day span, including 120 days of validation, to predict five-day outcomes in later windows. The image CNN trains separately in each window on 200 images for two epochs. Most of the eight test windows cover 85 days; the last covers 37, with gaps between windows. Each window starts in cash and liquidates at the end.
Research is separate from order execution. These simulations do not establish actual exchange fills, market impact or future performance.
Read the data and method notes ↗ · Read the numerical CNN source ↗
Back to market topics →