Thirty-one ways to predict crypto direction

The predictors

Every predictor is turned into a causal percentile rank within the trailing 365 days, centred on zero and multiplied by the sign the literature expects. A positive t-statistic therefore means the hypothesis holds. We regressed the forward 1-, 7- and 30-day return on each score with Newey–West standard errors, in two periods: up to 2022, and from January 2023 to September 2026.

What the predictors say

• Trend works. The 28-day and 12-week trends and the distance from the 200-day average had t-statistics between 3.4 and 4.2 on Bitcoin before 2022, and kept their sign afterwards.
• Contrarian indicators work backwards. High MVRV (market value to realised value), a high Puell multiple (miner revenue against its yearly average), extreme Wikipedia interest and a greedy Fear & Greed reading were all followed by further gains. They react to the price trend, so they inherit its momentum.
• Funding is not contrarian either. High funding was, if anything, followed by further gains on Ether (t −2.32 against the contrarian reading).
• Macro captured one regime. The financial conditions index had a t-statistic of 5.27 on Bitcoin before 2022, driven by the 2022 tightening, and −0.15 afterwards. Fed net liquidity, the dollar index, real yields, the VIX and the Nasdaq trend had no signal.
• Stablecoin flows were consistent across both assets and both periods, at a modest size, as described in the stablecoin post.

Selection and the model

The selection rule was fixed before looking at the post-2022 data: keep a predictor if |t| ≥ 2 on Bitcoin and |t| ≥ 1.5 with the same sign on Ether, both measured up to 2022, and use the sign that the data showed. Eleven of 31 passed:

For flipped predictors a negative t-statistic means the momentum reading held. After 2022, ten of the eleven kept their sign on Bitcoin, none with t ≥ 2 on its own. The same rule on the uncleaned stablecoin data selects the identical eleven.

The model averages the selected scores within each family and then takes the equal-weight mean of the five families (price, on-chain, attention, flows and macro). It goes long when the average is above +0.05, holds cash when it is below −0.05, and otherwise keeps its current position. A unit test checks that the production code matches the research code on more than 2,500 days.

Out of sample, 2023–2026

We also ran the model unchanged on nine altcoins that played no part in the selection (SOL, XRP, ADA, DOGE, LINK, LTC, AVAX, DOT and ATOM), using each coin’s own price and Bitcoin’s data for the other families. In 2023–2026 it had a better Sharpe than holding in 7 of 9 coins, a smaller drawdown in all 9 and a positive alpha in all 9, none of them individually significant. The median Sharpe was 0.53 against 0.34.

The model earns its keep by holding cash in crashes: it was out of the market for all of 2022, when Bitcoin fell 64%. In strong and choppy years it lags. On Bitcoin it trailed holding in 2023, 2024 and 2025, and in the sideways year 2025 it was whipsawed into a −23.5% loss against −6.3% for holding. Its better out-of-sample Sharpe comes from lower volatility and drawdowns, not from beating the market in any full year.

What did not help

• Machine learning on everything. A walk-forward ridge regression on all 31 predictors, refit monthly on past data only, had an out-of-sample Sharpe of 1.12 on Bitcoin and 0.22 on Ether, both worse than the simple average.
• Single families. Trend alone had a Sharpe of 1.11 on Bitcoin with a smaller drawdown (−29%). The stablecoin score traded as a continuous signal switched about 117 times a year and fees dominated.
• Volatility targeting. Scaling the position to a 40% volatility target raised the out-of-sample Sharpe to 1.27 on Bitcoin and 1.16 on Ether and cut the drawdowns to −28% and −37%. We did not deploy it, because the 40% level was chosen after seeing these results.
• Dropping the macro family would have lifted the Bitcoin Sharpe to 1.53. Doing that now would be fitting the model to its own test period, so the model stays as it was defined.

Caveats

• Short test period. Under four years out of sample, with no 2018- or 2022-style crash in it, and alphas that are not significant.
• Partly contaminated. The flows family contains the stablecoin signal, which had been examined on data that includes 2023–2026.
• Data vintage. Financial conditions and money supply are revised, and on-chain providers revise their attribution of exchange addresses. We used today’s vintages, not the values that were known at the time, so part of the pre-2022 strength of these predictors may be look-ahead.
• Overlapping families. The five family scores are correlated (0.2 to 0.7), and MVRV, Puell and attention largely echo the price trend.
• Costs. A flat 0.25% fee per side, with no spread or slippage and no interest on idle cash.

The model has run as a virtual shadow book next to the stablecoin bot since 29 September 2026; see the live results.