How I Built a 76% Win-Rate Bittensor Subnet Signal (Full Methodology)

Four days ago I set out to answer one question: can you systematically pick Bittensor subnets that will outperform over the next 30–60 days?

Everyone I asked had an opinion. Almost nobody had numbers.

So I built a backtest harness against a full year of on-chain data (2025-05-01 → 2026-04-21) and started testing. This post is the log.

Day 1 — The obvious strategies all scored 0.00

The first day I tested every strategy anyone would have told me to test:

Every rotation-based strategy got whipsawed. Every "stable" strategy underperformed. A year of Bittensor alpha markets has been too choppy for naive momentum or naive value to survive.

Lesson: Bittensor alphas don't trend the way stocks do. They spike and fade on a weekly timescale that destroys both momentum chasers and mean reverters.

Day 2 — Profiling actual winners

Rather than test more guesses, I flipped the problem: look at the subnets that actually won over the last year, and reverse-engineer the pattern.

The top-5 over May 2025 → April 2026:

• SN67: +1,088% — tao_in 380, price 0.001τ
• SN97: +2,825% — tao_in 700, price 0.002τ
• SN71: +1,304% — tao_in 492, price 0.0005τ
• SN52: +812%
• SN41: +623%

Notice anything? None of them had momentum when the gains started. They were small. Cheap. Under-discovered. By the time Twitter was talking about them, the move was already over.

The signal wasn't momentum. The signal was pre-momentum accumulation — capital flowing in before price had moved.

Day 3 — Building the filter

I translated that intuition into four concrete rules, tuned on the first 6 months of data and tested on the second 6 months:

1. tao_in in bottom 25% of the universe — small enough that real capital inflow moves the needle
2. Price in bottom 75% — not overpriced relative to the rest of the subnet economy
3. Emission above median — subnet is active, not dying
4. 14-day price change < 30% — hasn't already started pumping

Every subnet that passed all four filters at any given weekly snapshot became a "pick." Then I measured what happened to that pick over the next 30 days.

Result on held-out data: 67 wins out of 88 picks. 76% win rate.

(A "win" is defined as > 0% price return over 30 days — a conservative floor, since the real winners returned 3x–10x.)

Day 4 — Why it works (the mechanism)

Once I saw that the filter hit, I went back and profiled what those winning subnets had in common in terms of hidden dynamics. The pattern:

• Emission growing 21x faster than the universe median
• Stake growing 25x faster
• Volume growing 3.5x faster
• ...all while price lagged

That's a coiled spring. Capital is voting with actions — stake, emission, validator weight — while the tape hasn't caught up yet.

The filter doesn't "predict" anything. It identifies the gap between fundamentals and price.

The regime problem

Here's where most backtests lie: they don't tell you when the strategy fails.

Mine fails when established subnets rally. When SN1 Apex or SN3 are ripping +30%, capital stays parked there instead of rotating into speculative small subnets. The filter fires — but the picks don't go anywhere.

So the live version has a regime filter:

• Active when SN1's 30-day return is below +5%
• AND the universe-aggregate 30-day return is below +10%
• Otherwise: no picks published that week

This makes the strategy fewer picks, higher conviction. Which is the right tradeoff when you're charging subscribers for a signal.

What I'm not hiding

1. 1 year of data is not 10 years. The filter survives the split I tested it on, but Bittensor dTao itself is young. The market structure has changed visibly since late 2025.
2. Dedup risk exists. If every SN88 miner runs a similar logic, the mining scores collapse via SN88's similarity penalty. I'm watching this live.
3. The winning subnets are real winners. But so are the losing subnets. 21 of 88 picks lost money. The 76% is a floor, not a guarantee.
4. Percentile thresholds beat fixed thresholds. The universe changes — what was "bottom 25% tao_in" in May 2025 ($100K) was different by Feb 2026 ($500K). The filter uses percentiles so it re-anchors.

Live deployment

As of April 22, 2026, this strategy is deployed on Bittensor SN88 UID 180 with 257 alpha staked. Same subnets, same weights you see on the public track record. Same scoring rules SN88 uses to rank miners.

If it ranks top-20 on SN88 in the next 30 days, the strategy survived its first contact with reality.

If it doesn't, I publish why.

Where to find it

• Public track record → signals.subnetaiq.io/signals/track-record — every historical pick, current regime, win/loss breakdown
• Current regime → signals.subnetaiq.io/signals — free to check any time
• This week's portfolio → SubnetAIQ Validator subscribers ($9.99/mo)

If you're picking subnets by gut, this might save you some research cycles. If you're skeptical — read the track record, pick it apart, and tell me where it's wrong.

— SubnetAIQ