How the Signals Work

Every input in the system earns its place by surviving the same test: does it actually have predictive value on data it never saw during development? If not, it gets cut. Here's what currently makes the cut and where each one comes from.

The on-chain inputs — MVRV, NUPL, and related metrics — are pulled daily from Santiment's enterprise data feed, which reads directly from the Bitcoin blockchain. Price data lives in our own dedicated database, updated continuously. Fear & Greed and ETF flow data are pulled from their respective sources and ingested every morning before the signal run. By the time the daily report fires, everything is current and the analysis is running against live data.

We tested every combination of these inputs against 10+ years of price history under strict rules: the data was split so no signal could be optimized on the same data it was tested against, transaction costs were included, and sample sizes had to be large enough to carry statistical weight. Results were tiered by win rate and Sharpe ratio — how good the return was relative to the risk taken.

The signal work is not just about what we validated. It is equally about what we threw out.

We also rejected: any signal with fewer than 20 test occurrences, any signal showing a Sharpe above 5.0 in testing (unrealistically good — almost certainly an artifact of small sample size), combinations of four or more signals simultaneously, and order book signals after extensive testing produced no viable edge after transaction costs.

The inputs you see today are not the inputs we started with, and they are not necessarily the final picture. We have run over 300 experiments building and refining the scoring model. One significant finding from that work: not all inputs should be weighted equally. Shorter-memory signals like RSI and Fear & Greed benefit from a different lookback window than slower-moving signals like MVRV and drawdown. That architectural discovery improved out-of-sample performance meaningfully and is now built into the production system.

We continue to evaluate new inputs — volatility patterns, options market data, institutional positioning data, and others. Some will pass the test. Most won't. The ones that do will be added. The ones already in the system that show signs of degrading as market structure evolves will be demoted or removed. The signal system is a living research program, not a fixed model.

Even the best entry signal means something different depending on where in the cycle it fires. A Silver-tier signal at a confirmed cycle bottom is a strong entry. The same signal in the middle of a bear market may just be a temporary rally before more downside. Buying a bear market rally thinking it's a bottom is one of the most expensive mistakes in this asset class.

The Markov cycle model exists to answer that second question: where in the cycle are we, and what does history say happens next from here?

Bitcoin has a well-documented four-year market cycle, loosely anchored to its supply halving events. We defined four distinct phases using MVRV ratio as the objective measure. The data defines the phases — not an analyst's opinion.

We then analyzed over 4,800 days of Bitcoin price history going back to 2013. For every day in that history we recorded: what phase were we in, and what phase did we transition to over the following 30, 60, and 90 days? After building that complete transition map from four full Bitcoin cycles of actual data, we can now ask a precise question: given today's phase and MVRV reading, what does the historical record say the probability distribution of outcomes looks like at 30, 60, and 90 days forward?

The same discipline that governs signal validation applies here. We used a walk-forward validation approach: the transition matrices were trained on the earliest portion of Bitcoin's history, then tested against how well those probabilities matched what actually happened in periods the model never saw. The model passed.

We tested eight different configurations — varying how we defined accumulation phase boundaries, varying how we binned the MVRV zones — and ranked them all by how well their predicted probability distributions matched real historical outcomes. The best-performing configuration is what runs in production. As more cycle data accumulates, the transition probabilities will be updated. More cycles means more confident estimates.

Every morning at 10:05 UTC, the model reads the current MVRV, confirms the phase, runs 10,000 simulations of what history says happens next, and produces a forward probability distribution. The report shows you:

The signals tell you whether conditions look like a bottom. The Markov model tells you where in the cycle you likely are. Together, they answer the question every investor actually has: should I be buying here, and how much?

This is not a crystal ball. No model predicts the future with certainty. What these tools do is put the weight of evidence on your side. You will not catch every bottom perfectly. You will occasionally enter into what turns out to be a bear rally.

But across a sufficient number of cycles, acting on validated high-tier signals in cycle phases where the Markov model confirms elevated probability of accumulation has produced outcomes that substantially outperform undisciplined buying. Most investors lose money in this asset class not because good opportunities don't exist, but because they act emotionally at exactly the wrong moments. These tools exist to make that less likely.