I built a cross-sectional momentum strategy using on-chain capital flows across 50 blockchains. The signal had a positive Information Coefficient\u200a\u2014\u200abut the portfolio lost money. Here\u2019s how that\u2019s possible, and what it teaches about crypto\u00a0alpha.<\/em><\/p>\n TVL (Total Value Locked) growth into a blockchain should predict its native token appreciating\u200a\u2014\u200amore capital = more users = higher returns. I tested this across 50 chains over 4\u00a0yearsThe signal has a positive IC of 0.043 (rank correlation between TVL flow and forward returns), but a t-statistic of only 1.62\u200a\u2014\u200abelow the 2.0 significance thresholdCritical finding: the portfolio returns are NEGATIVE. Going long high-inflow chains and short outflow chains produces a Sharpe of -0.31. The signal works in reverse<\/em> at the portfolio levelRoot cause: TVL flow is just crypto price momentum in disguise. When I stripped out the price effect (residual flow), the IC collapsed to zero. There is no genuine capital flow\u00a0alphaVerdict: Not tradeable. TVL flow momentum = repackaged price momentum, and crypto price momentum mean-reverts at monthly\u00a0horizons<\/p>\n In traditional finance, capital flows predict returns. When international fund managers pour money into a country\u2019s stock market, that market tends to outperform. The mechanism is straightforward: inflows push prices up, and institutional investors tend to move capital toward markets with genuinely better prospects.<\/p>\n I wanted to test whether the same logic works in crypto, using Total Value Locked (TVL) as the flow\u00a0measure.<\/p>\n TVL measures the total capital deployed in a blockchain\u2019s DeFi ecosystem\u200a\u2014\u200athe money sitting in lending protocols, DEXes, liquid staking, yield farms. When TVL grows on Solana, for instance, it means capital is flowing into that ecosystem. Someone, somewhere, decided Solana\u2019s DeFi opportunities were worth locking up capital\u00a0for.<\/p>\n The hypothesis: Chains experiencing TVL growth (capital inflows) should see their native tokens appreciate. This creates a tradeable cross-sectional momentum signal\u200a\u2014\u200ago long the chains attracting capital, go short the chains losing\u00a0it.<\/p>\n The economic reasoning was compelling:<\/p>\n TVL growth = \u201csmart money\u201d voting with their\u00a0walletsPositive feedback loop: more TVL \u2192 more liquidity \u2192 better yields \u2192 more users \u2192 token appreciationMarkets should underreact to capital flow momentum (behavioral underreaction)Academic research documents cross-cryptocurrency return predictability<\/p>\n But I had one concern from the start: TVL mechanically includes token price changes. If SOL doubles in price, Solana\u2019s TVL roughly doubles too, even if no new capital entered. This could mean TVL flow isn\u2019t measuring real capital flows at all\u200a\u2014\u200ajust repackaging price momentum.<\/p>\n I designed a specific test for this (Signal 4 below). The answer would determine whether this was genuine alpha or an artifact.<\/p>\n The data came from DefiLlama\u2019s free API\u200a\u2014\u200ano $300\/month Pro subscription needed:<\/p>\n Chain TVL history: \/v2\/historicalChainTvl\/{chain}\u200a\u2014\u200adaily TVL for 300+ chains back to\u00a02020Chain metadata: chains.json\u200a\u2014\u200amaps each chain to its CoinGecko token\u00a0IDToken prices: coins.llama.fi\/chart\/coingecko:{id}\u200a\u2014\u200adaily OHLCV for chain\u00a0tokens<\/p>\n I built a 97-chain panel with 80,284 daily observations spanning January 2020 to February 2026. After applying tradability filters (TVL > $30M median, 12+ months of history, starting January 2022), the investment universe narrowed to 50 chains per month on average\u200a\u2014\u200aincluding ETH, SOL, BNB, AVAX, ARB, OP, SUI, TRX, and\u00a0others.<\/p>\n Figure 1: Strategy parameters. Monthly rebalance with quintile sort across 50 blockchain chains.<\/em><\/p>\n I designed four signal variants to test the thesis from different angles:<\/p>\n Figure 2: The four TVL flow signals. Signal 4 (residual flow) is the intellectual litmus test\u200a\u2014\u200aif it works, we have genuine alpha. If not, we\u2019re just repackaging price momentum.<\/em><\/p>\n The key insight in this design is Signal 4: Residual Flow. Every month, I run a cross-sectional regression:<\/p>\n TVL_change(%) = alpha + beta1 * token_price_change(%) + beta2 * BTC_change(%) + epsilon<\/p>\n The residual (epsilon) captures TVL changes that can\u2019t be explained by token price movements<\/em>. If someone deposits 1,000 ETH into Arbitrum\u2019s DeFi protocols, that\u2019s a genuine capital inflow that should show up as positive residual flow\u200a\u2014\u200aregardless of what ETH\u2019s price did that\u00a0month.<\/p>\n If residual flow IC > raw flow IC \u2192 genuine capital flow alpha (thesis validated) If residual flow IC \u2248 0 \u2192 just repackaged price momentum (thesis\u00a0dead)<\/p>\n Monthly rebalance, quintile sort on signal\u00a0valuesLong Q1 (strongest TVL inflows), Short Q5 (weakest\/outflows)BTC single-factor beta hedge (neutralize market exposure)Tiered transaction costs: 10 bps (large caps like ETH, SOL, BNB), 20 bps (mid caps), 40 bps (small\u00a0caps)<\/p>\n I tested all 16 combinations (4 signals \u00d7 4 forward return horizons):<\/p>\n Figure 3: Information Coefficient heatmap. Green = positive predictive power. The raw TVL signals show modest positive ICs (0.03\u20130.05), but all t-statistics fall below 2.0. The residual flow row (bottom) is\u00a0dead.<\/em><\/p>\n Figure 4: Top signal-horizon combinations ranked by t-statistic. Note how residual flow variants are marked \u201cDEAD\u201d\u200a\u2014\u200athis is the smoking\u00a0gun.<\/em><\/p>\n The best combination\u200a\u2014\u200atvl_flow_30d at a 30-day horizon\u200a\u2014\u200aproduced an IC of 0.043 with a t-statistic of 1.62. In plain English: there’s a modest positive rank correlation between TVL flow and forward 30-day returns, but it’s not statistically significant at conventional thresholds.<\/p>\n More revealing: the residual flow signal is effectively zero (IC = 0.010, t = 0.36). This means once you strip out the mechanical price effect, TVL changes have zero predictive power for token\u00a0returns.<\/p>\n Figure 5: Annual IC for tvl_flow_30d. The signal has strengthened recently (0.099 in 2025, 0.195 in early 2026), but the overall t-statistic across all years is still only 1.62\u200a\u2014\u200ainsufficient evidence that this isn\u2019t\u00a0noise.<\/em><\/p>\n An interesting pattern: the IC was near zero in 2022\u20132023 but strengthened in 2024\u20132025. This could mean the signal is emerging, or it could mean we\u2019re seeing recent random variation in a small sample. Without a t-stat above 2.0 over the full period, I can\u2019t distinguish signal from\u00a0noise.<\/p>\n This is the most important chart in this post. It shows why a positive IC doesn\u2019t guarantee a profitable portfolio.<\/p>\n Figure 6: Average monthly returns by TVL flow quintile. Q1 = chains with the largest TVL outflows, Q5 = chains with the largest inflows. The thesis predicts Q5 > Q1. Reality: Q1 crushes\u00a0Q5.<\/em><\/p>\n The signal works in complete reverse at the portfolio level:<\/p>\n Q1 (TVL outflows): +0.36%\/month\u200a\u2014\u200aoversold chains bounce\u00a0back<\/em>Q5 (TVL inflows): -1.12%\/month\u200a\u2014\u200acrowded chains underperform<\/em>Long\/Short spread: -1.48%\/month, Sharpe =\u00a0-0.31<\/p>\n Going long the chains everyone is piling into and shorting the chains people are abandoning loses money consistently<\/em>. The contrarian trade would have\u00a0worked.<\/p>\n This is a genuine statistical paradox that I think is worth explaining, since it can trap researchers who only look at\u00a0IC:<\/p>\n The IC measures rank correlation across the entire cross-section. It asks: \u201cDoes higher signal generally predict higher returns?\u201d And the answer is a weak yes\u00a0(0.043).<\/p>\n But the portfolio only captures the extremes<\/em> (Q1 vs Q5). The quintile returns are non-monotonic\u200a\u2014\u200athe relationship between signal and returns breaks down at the tails. Middle quintiles might follow the expected pattern while the extremes\u00a0invert.<\/p>\n Think of it this way: chains with moderate TVL inflows might slightly outperform chains with moderate outflows (contributing to a positive IC), but chains with extreme<\/em> inflows (crowded trades) underperform chains with extreme<\/em> outflows (oversold bounces). The IC averages over the whole distribution, but the L\/S portfolio lives at the extremes.<\/p>\n Lesson: Always compute quintile spread returns. IC alone is not sufficient.<\/p>\n Figure 7: The key diagnostic. Blue bars show raw TVL flow IC (positive but weak). Red bars show residual flow IC (near zero everywhere). Once you remove the token price effect, there\u2019s nothing\u00a0left.<\/em><\/p>\n This chart tells the whole story. The raw TVL flow signal has modest predictive power (blue bars above zero at 30d, 60d, 90d horizons). But the residual flow\u200a\u2014\u200aTVL changes after accounting for token price movements\u200a\u2014\u200ahas essentially zero IC at every\u00a0horizon.<\/p>\n What this\u00a0means:<\/p>\n TVL flow is measuring price momentum<\/em>, not genuine capital\u00a0flowsWhen SOL goes up 30%, Solana\u2019s TVL mechanically rises ~30%\u00a0tooThe \u201cTVL inflow\u201d signal is really saying \u201cthis token\u2019s price went up recently\u201dCrypto price momentum mean-reverts<\/em> at monthly horizons (well-documented)Therefore: buying \u201chigh TVL flow\u201d tokens = buying recent winners = buying into mean-reversion<\/p>\n This explains the quintile reversal perfectly. The signal identifies recent momentum (positive IC at the rank level), but crypto momentum reverses at 30-day+ horizons, so the extreme winners (Q5) subsequently underperform while the extreme losers (Q1) bounce\u00a0back.<\/p>\n I use a formal 4-gate validation framework before proceeding to backtesting. The strategy must pass at least 3 of 4\u00a0gates:<\/p>\n Figure 8: Decision gate results. Only 1 of 4 gates passed. This triggers a HARD STOP\u200a\u2014\u200ano point running a backtest on a dead\u00a0signal.<\/em><\/p>\n 1 out of 4 gates passed. Hard\u00a0stop.<\/p>\n I could have spent another week building a walk-forward backtest engine, testing 48 robustness configurations, and computing Deflated Sharpe Ratios. But the signal validation told me everything I needed to know in 30 minutes of compute\u00a0time.<\/p>\n This is the value of decision gates: they prevent you from sunk-cost-fallacy-ing your way through a doomed\u00a0project.<\/p>\n This is the biggest trap in on-chain data. TVL = quantity_locked \u00d7 token_price. When ETH goes up 50%, Ethereum\u2019s TVL goes up ~50% even if zero new capital entered. Any signal built on raw TVL changes is contaminated by price momentum.<\/p>\n Fix: Always compute residual flow (TVL change after regressing out price effect) before making claims about \u201ccapital\u00a0flows.\u201d<\/p>\n This is well-documented but bears repeating. In equities, momentum works at 2\u201312 month horizons. In crypto, momentum works at 1\u20137 day<\/em> horizons and then reverses. Building a monthly-rebalance momentum strategy in crypto is swimming against the\u00a0current.<\/p>\n Implication: If you find a crypto signal with monthly IC > 0, check whether it\u2019s just disguised momentum. If so, the portfolio will likely show reversal at execution frequency.<\/p>\n Spearman IC measures average rank correlation across the entire cross-section. A positive IC of 0.04 means \u201chigher signal \u2248 slightly higher returns on average.\u201d But portfolios live in the tails. If the extremes don\u2019t follow the overall pattern, the L\/S portfolio can lose money despite positive\u00a0IC.<\/p>\n Fix: Always compute quintile spreads and check monotonicity before backtesting. IC is necessary but not sufficient.<\/p>\n In a field full of repackaged factors, the fastest way to check novelty is to residualize your signal against known factors. If the residual IC collapses to zero, you haven\u2019t found anything new\u200a\u2014\u200ayou\u2019ve just found a noisy proxy for an existing\u00a0factor.<\/p>\n I built this test into the experiment design from Day 1, and it saved me from a false positive.<\/p>\n The previous project (DeFi Revenue Value) ran 5 full experiments before discovering the signal was fragile. This time, I applied a lesson from that failure: test quarterly IC early, validate gates strictly, and stop when the evidence says\u00a0stop.<\/p>\n Result: 3 experiments instead of 5. Days of compute time saved. The rejection was clear, unambiguous, and well-documented.<\/p>\n Thesis: REJECTED.<\/p>\n TVL flow momentum is not a viable cross-sectional alpha source for blockchain tokens. The signal is mechanically dominated by price momentum, which mean-reverts at the monthly rebalance frequency. The residual flow analysis confirms there is no genuine capital flow alpha buried underneath.<\/p>\n Possibly:<\/p>\n Contrarian (reverse the signal): Going long<\/em> TVL outflow chains and short<\/em> inflow chains would have been profitable. But this is a different thesis (mean reversion, not momentum), and the negative Sharpe on the forward direction (-0.31) suggests weak magnitude even for the\u00a0reverseHigher frequency (daily\/weekly): TVL flow at very short horizons, before the momentum reversal kicks in, might capture genuine short-term price pressure. But the data granularity and execution costs make this challenging for cross-sectional chain\u00a0tokensTVL composition changes: Instead of total TVL flow, tracking which protocols<\/em> are gaining\/losing share within a chain might capture genuine innovation momentum. This is a different signal\u00a0entirely<\/p>\n While the trading signal failed, the project created reusable infrastructure:<\/p>\n Chain TVL pipeline: 300 chains, free API, incremental refreshChain-to-token mapping: 300 CoinGecko IDs mapped to blockchain namesPaginated price API: 3-window approach for coins.llama.fi covering 2020\u20132026Residual decomposition framework: Cross-sectional regression separating price effect from genuine\u00a0flows<\/p>\n These assets are ready for the next hypothesis.<\/p>\n Date: February\u00a02026Period: January 2022\u200a\u2014\u200aFebruary 2026 (4\u00a0years)Universe: 50 blockchain chains with TVL > $30M and tradeable native\u00a0tokensData Sources: DefiLlama (chain TVL, free API), CoinGecko via coins.llama.fi (token\u00a0prices)Signals tested: 4 TVL flow variants \u00d7 4 horizons = 16 combinationsMethodology: Cross-sectional Spearman IC, quintile sort, L\/S portfolio estimation<\/p>\n Disclaimer: This research is for educational purposes only. Past performance does not guarantee future results. Always do your own due diligence before making investment decisions.<\/em><\/p>\n Tags: #QuantitativeFinance #Crypto #DeFi #TVL #CrossSectionalMomentum #AlphaResearch #FailedStrategies<\/p>\n I Tested Whether \u201cFollow the Money\u201d Works in Crypto. It Doesn\u2019t \u2014 The Signal Runs Backwards.<\/a> was originally published in Coinmonks<\/a> on Medium, where people are continuing the conversation by highlighting and responding to this story.<\/p>","protected":false},"excerpt":{"rendered":" How TVL Inflows Into Blockchain Ecosystems Predict Token Returns (in the Wrong Direction) I built a cross-sectional momentum strategy using on-chain capital flows across 50 blockchains. The signal had a positive Information Coefficient\u200a\u2014\u200abut the portfolio lost money. Here\u2019s how that\u2019s possible, and what it teaches about crypto\u00a0alpha. TL;DR TVL (Total Value Locked) growth into a […]<\/p>\n","protected":false},"author":0,"featured_media":141022,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-141021","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-interesting"],"_links":{"self":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/posts\/141021"}],"collection":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=141021"}],"version-history":[{"count":0,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/posts\/141021\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=\/wp\/v2\/media\/141022"}],"wp:attachment":[{"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=141021"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=141021"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mycryptomania.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=141021"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}TL;DR<\/h3>\n
Part 1: The Hypothesis\u200a\u2014\u200aFollow the\u00a0Money<\/h3>\n
Part 2: Data & Methodology<\/h3>\n
Data Sources (All\u00a0Free)<\/h3>\n
Signal Design: 4\u00a0Variants<\/h3>\n
Portfolio Construction<\/h3>\n
Part 3: Statistical Analysis\u200a\u2014\u200aThe Numbers Look Promising\u2026 Then\u00a0Don\u2019t<\/h3>\n
IC Matrix: Weak But\u00a0Positive<\/h3>\n
The IC Evolves Over\u00a0Time<\/h3>\n
Part 4: The Reversal\u200a\u2014\u200aWhen Positive IC Produces Negative\u00a0Returns<\/h3>\n
How Is a Positive IC Possible With Negative Portfolio Returns?<\/h3>\n
Part 5: The Litmus Test\u200a\u2014\u200aIs This Just Price Momentum?<\/h3>\n
Part 6: Decision Gate Validation<\/h3>\n
Part 7: Key Lessons\u00a0Learned<\/h3>\n
1. TVL Mechanically Includes Token Prices\u200a\u2014\u200aAlways Decompose<\/h3>\n
2. Crypto Momentum Mean-Reverts at Monthly\u00a0Horizons<\/h3>\n
3. Positive IC Does Not Guarantee Positive Portfolio Returns (IC\u00a0Paradox)<\/h3>\n
4. Residual Signal Design Is Your Best Friend for Novelty Verification<\/h3>\n
5. Early Validation Saves Enormous\u00a0Time<\/h3>\n
Part 8: Final\u00a0Verdict<\/h3>\n
Could a Modified Version\u00a0Work?<\/h3>\n
What This Research\u00a0Produced<\/h3>\n
About This\u00a0Research<\/h3>\n