Research · extended study
Working paper · under reviewNo Edge Without Information
An empirical study of tradeability in decentralized-exchange-only cryptocurrencies
This study makes a single argument. In decentralized-exchange markets of small, mostly non-CEX-listed tokens, an exploitable edge cannot be built from price information alone. Price-based timing fails, the failure is structural rather than methodological, the non-timing strategies a practitioner would reach for next fail too, and the only signal that beats a null is non-price and merely predicts crashes. The pieces below let you reproduce the core mechanics in the browser.
Data partners
Their on-chain and market data made this study possible. A substantially expanded version 2, with broader coverage and deeper analysis thanks to their support, is in progress.
View on SSRN
working paper, not yet posted
Code & engines
github.com/DaruFinance/dex-tradeability-study
The result in three lines
4,990 pairs · 27 chains
No price-based edge
A long-only, realistically-costed trader has no exploitable edge from price information at any horizon. Real data performs worse than its own bar-shuffled null.
fifteen analyses
The failure is structural
It survives the intrabar exit convention, the selection regime, and the cost model, and it extends to hedging, liquidity provision, arbitrage, regime timing, portfolios, and order flow.
one exception
Skill in the left tail
A cross-sectional learner beats its null, but it predicts which coins crash, not which rise. A long-only trader cannot monetise it.
Overview
DEX-only tokens are the part of the crypto market least exposed to professional high-frequency competition, which makes them a natural place to ask whether an ordinary participant can extract systematic profit. Two features make the question sharp: on an automated market maker there is no native short, so directional edge can only come from the long side; and the per-trade cost is large and endogenous to size, routinely 150 to 450 basis points round-trip.
We assemble a survivorship-aware corpus of 4,990 trading pairs across 27 blockchains and run the evidence in five movements, each held to the same discipline: a bar-shuffled null control and full-distribution reporting. Price-based timing shows no detectable edge at any horizon, and real data fails to beat even its own (inflated) shuffle, on a return process that is negative-drift, heavy-tailed, volatility-clustered, and anti-momentum. That negative survives the obvious objections, and it extends to every non-timing strategy a practitioner would try next.
Interactive · cost frontier
Drag pool depth and trade size. Watch fee, slippage, and gas trade off into a U-shaped round-trip cost. On a deep pool the floor sits near 160 bp; gross timing edge is ~0, so net return is roughly minus this.
study default sizing (0.25% of reserve) = $1k → 217 bp
swap fee
60
bp
slippage
45
bp
gas
127
bp
round-trip cost
232 bp
below the ~300 bp tradeability line
x: trade size (log) · y: round-trip cost (bp). Left arm = gas-dominated, right arm = slippage-dominated. The minimum is the cheapest you can trade this pool.
The cost wall
Edge has to clear cost before anything else. On an AMM the round-trip cost is not a fixed number; it is a U-shaped function of trade size, dominated by fixed gas on small trades and by constant-product slippage on large ones. The minimum, near the study's 0.25%-of-reserve sizing, sits around 160 basis points on a deep pool and far higher on a thin one. Drag the pool depth and trade size and watch the three components trade off.
Interactive · null control
A long-only momentum strategy on a simulated DEX market vs. the same strategy on a bar-shuffled copy. If real beats noise, real should win. It does not.
REAL data
0%
strategies with PF > 1
0.27
median PF
-3.53%
mean net/trade
NULL (bar-shuffled)
0%
strategies with PF > 1
0.38
median PF
-2.65%
mean net/trade
→ With these knobs the structure is no longer adverse; raise anti-momentum or lower drift.
Worse than noise
The load-bearing finding is not merely that timing fails, but that real data performs worse than a bar-shuffled copy of itself. The shuffle keeps each coin's return distribution and destroys only the order of the bars. Because the order is what hurts a long position, clustered crashes on a downward, mean-reverting drift, shuffling it away helps. The demo runs a long-only momentum rule on a simulated market and on its own shuffle; the shuffle wins.
The same pattern recurs across the alternative-channel studies, where several apparent edges turn out to be the shuffle paying the per-fill cost on destroyed structure, not alpha.
Interactive · crash avoidance
Coins sorted into deciles by the ranker's predicted score. Raise the model skill and watch the deciles fan out, yet every one stays below zero. The skill is in the left tail.
Long-only, the top decile is still a loser. The model knows what falls, not what rises.
The one signal, and why it does not pay
A single test beats its null: a leakage-free cross-sectional gradient-boosted ranker on point-in-time causal features. Its out-of-sample rank information coefficient is small but real at the 7- and 14-day horizons (about 0.06 to 0.08, label-permutation p around 0.005) and insignificant at 30 days. But its skill is crash avoidance, not return seeking. Sort the universe into deciles by the model's score and every decile's median forward return is negative; the top predicted decile is the least bad, not the best. The information is in the left tail, which a long-only AMM trader cannot reach, so we report it as a lead, not an edge.
Method
- Long-only on the AMM (no native short), with a per-fill round-trip cost on both legs: swap fee plus constant-product slippage plus gas.
- Causal, walk-forward evaluation: parameters chosen in-sample, scored out-of-sample, with intrabar take-profit/stop using the bar high and low.
- A mandatory bar-shuffled null control: every claimed effect is re-run on a surrogate that destroys temporal order while keeping each coin's return distribution. An effect is real only if it beats the null net of cost.
- Full-distribution reporting (medians and the whole spread of outcomes), never a hand-picked winner, because the returns are extremely fat-tailed.
- Survivorship-aware throughout: the universe is sourced from currently-live pools, so every cross-sectional and holding metric is read as an upper bound.
Reproducibility
The analysis engines and the source of the write-up are collected in a companion repository, dex-tradeability-study. The interactive figures on this page are self-contained and seeded, so the mechanics they illustrate reproduce exactly on every load.
Cite
See also
A companion methodological study on permutation testing and forward selection is at Research, and the narrative note The edge is in the process develops the selection-discipline theme that runs through both.