Tao-Quant paper trading · live

A quant paper-trading sandbox for Bittensor subnet alpha tokens — built to learn what works before any real capital is at risk.

What this system does

Tao-Quant runs a small portfolio of strategies against the live Bittensor market every 30 minutes. Every trade is paper — no real money. Its job is to generate clean, structured data about which signals actually predict subnet returns so we can graduate the winners to real capital later.

How the pieces fit

Architecture: API -> Snapshot Writer -> Screener -> Scorers -> Portfolios -> Execution -> DB -> Dashboard

Five sleeves run in parallel

Sleeve	What it trades	Why we run it
`subnet_alphanomics`	Top-6 subnets by 9-component composite score	The algorithmic research strategy — selects from the full universe by score.
`subnet_liquidity`	Top-6 subnets by pool depth	Dumb baseline — proves whether scoring beats just buying the biggest.
`subnet_dsv`	DSV Fund's broad 20-name allowlist, equal weight	Curated universe — does fund-grade screening beat algorithmic selection?
`subnet_lewis`	Lewis Smith / Mark Jeffrey 7-position weighted allocation, 50% root staking	Conservative reference allocation from the public 10x Challenge dashboard.
`subnet_kidd`	Siam Kidd / DSV Fund 6-position concentrated allocation, 30% root staking	Aggressive, revenue-screened. CIO of world's first TAO-exclusive hedge fund.

Why five sleeves? Each is a separately-funded $10K paper sleeve so equity curves are directly comparable. liquidity is the dumb-money floor; alphanomics tests whether algorithmic scoring adds value; dsv tests broad curated universe vs algorithm; lewis and kidd are real published portfolio allocations from the jackson-video-resources/bittensor-investing-agent repo (Lewis Smith's 10x Challenge). The two weighted sleeves include SN0 root staking, which the others don't — see Methodology.

Key facts

Cadence: 30-minute ticks via Windows Task Scheduler. Snapshots write every tick; trades fire only when a position drifts > 5% of sleeve equity from target.
Universe: 129 Bittensor subnets, screened down to ~30 eligible names per tick.
Capital: $10,000 paper per sleeve. Real-money capacity ceiling is around $10K total per Maymin (2026) — see Methodology tab.
Data: taostats.io free tier (5 req/min, throttled to 12.5s between calls). Coingecko for spot TAO/USD.
Friction modeled: per-pool fee_rate from API + AMM slippage (trade_tao / pool_tao × 0.5, capped 5%).

Bittensor, TAO, and dTAO subnets — the short version

Bittensor is a decentralized network where independent “subnets” compete to perform useful AI work (inference, training, scraping, audio, bio, etc). Each subnet has its own token (an “alpha” token). TAO is the macro asset that flows between them.

The three layers

Three layers: TAO reserve currency, 129 specialized subnets, one dTAO AMM pool per subnet

Why subnet prices move

Emissions allocation. The protocol allocates new TAO emissions to subnets weighted by a 30-day half-life EMA of their net stake flow (“Taoflow”). More inflows → more emissions → buying pressure.
AMM mechanics. Each subnet has a single constant-product pool (TAO ⇄ alpha). A buy of x TAO into a pool with reserve τ produces a 2·x/τ % price change to first order — meaning smaller pools amplify more. This is Maymin (2026) Proposition 1.
Real-world demand. Validators must hold the subnet’s alpha token to vote on miner quality, so subnets with healthy validator participation have organic floor demand.
Network effects. A subnet with active miners + validators + rising emissions creates a flywheel; one without these collapses.

How a trade actually works

Trade sequence: trader sends TAO, pool applies constant product, returns alpha tokens at new price, balance recorded on-chain

Why the SMB factor exists

Maymin’s arXiv 2603.29751 paper studied the cross-section of subnet returns and found that a small-minus-big portfolio (long the smallest market-cap subnets, short the largest) earned ~1% per day with a gross Sharpe of 3.84. The mechanism is the AMM: identical TAO emissions cause much bigger % moves in small pools, mechanically.

The catch — capacity. The same AMM that amplifies small-pool returns also amplifies your own trade impact. The strategy capacity ceiling per Maymin is roughly $10K — at $100K AUM you’re paying so much slippage you eat the alpha. This is a research and small-capital strategy, not a scalable one.

Glossary

Term	What it is
`TAO`	Native asset of the Bittensor network. Roughly analogous to ETH on Ethereum.
`Subnet`	A specialized network within Bittensor that does one type of work (e.g. text inference). Each has a number (netuid) and a name.
`Alpha token`	A subnet’s own token. Trades against TAO via the subnet’s AMM pool.
`dTAO`	The decentralized-TAO upgrade that gave each subnet its own pool and price discovery (vs the prior proportional-emission model).
`Taoflow`	The 30-day EMA of net stake flow into a subnet. The protocol uses this to set emissions.
`Validator`	An entity that scores miner work and earns emissions. Must stake TAO + hold subnet alpha.
`Miner`	An entity that performs the subnet’s actual work (inference, scraping, etc) and earns emissions.
`SMB factor`	Small-minus-big — long small market caps, short large. The dominant alpha factor per Maymin.

How the bot decides what to trade

Every 30 minutes the bot pulls a live snapshot, runs each subnet through a screener, scores the survivors, builds a target portfolio, and trades the deltas. Each step is intentionally simple so we can attribute outcomes to specific signals.

The full pipeline

Full pipeline: 30-minute tick, pull data, screen, score both sleeves, build top-6, compare to current, trade or no-op, mark equity, write DB

1. The screener — drop dust and unhealthy subnets

Out of 129 subnets, we eliminate any that fail any of the following gates:

Filter	Threshold	Reason
Liquidity (raw)	≥ 1e11 (~100 TAO)	Drops dust pools where price is meaningless.
Pool depth ceiling	≤ 50,000 TAO	Pools above this have negligible AMM amplification on a $10K trade (Maymin/DSV bound).
Active validators	≥ 8	Network health — too few validators = capture risk.
Active miners	≥ 10	Real activity — ~50% of subnets have ≤1 miner.
1-week price change	≥ −30%	No falling-knife catches.
Price > 0	strict	Sanity.

2. The alphanomics scorer (v4)

Each surviving subnet gets nine component scores in [0,1]. The composite is a weighted sum:

Component	Weight	What it measures	Source
size_inverse	0.25	Percentile rank by −mcap — small wins	Maymin SMB
alphanomics_short	0.15	1-day net stake flow percentile	Taoflow proxy
alphanomics_med	0.15	7-day net stake flow percentile	Taoflow proxy
momentum_1m	0.15	30-day price change, clipped ±50%	Maymin WML30
momentum_1w	0.10	7-day price change, clipped ±20%	Maymin WML7
emission_yield	0.05	Daily emission / mcap	Maymin HML_EMIS
buy_sell_pressure	0.05	24h buy_vol / total_vol	Order-flow
network_health	0.05	validators × miners, percentile	Sanity tilt
flow_acceleration	0.05	1d flow vs 7d daily-avg	Catalyst

Why size dominates the weights. The Maymin paper found SMB has Fama-MacBeth t = 3.23 vs < 2 for most other factors in the cross-section. The 0.25 weight reflects “strongest signal in the academic record + mechanically grounded in the AMM.” Other components are present mostly to diversify against pure size, not because we trust them as much.

3. Portfolio construction

Top-K = 6. Six equally-weighted positions per sleeve. Empirically this gives enough diversification without spreading capital so thin that AMM slippage dominates each trade.
Max 25% per name. Hard cap so a single subnet can never blow up the sleeve.
Equal weighting, not score-weighted. Score-weighted concentrates in the top name and amplifies single-subnet risk.

4. Execution & friction

Trade trigger: only when |target − current| > 5% of sleeve equity. Acts as the natural friction filter to avoid over-churning.
Slippage: min(0.05, (trade_tao / pool_tao) × 0.5) — capped at 5%. Linear approximation of AMM impact.
Fees: per-pool fee_rate from the API, ~5 bps default.
Mark-to-market: equity is stamped every tick regardless of whether a trade fires, so we get a continuous P&L curve.

5. The liquidity baseline (control arm)

Same screener, same K=6 equal-weight construction, same execution — but the score is just raw pool liquidity. This is the dumb-money strategy. If alphanomics can’t beat it, the composite score is adding nothing.

6. The DSV Fund picks (curated-allowlist arm)

DSV Fund publishes a list of 20 subnets they have meaningful capital in. We treat that list as a hard allowlist: same screener, same execution, but the score becomes a binary “is it on the list, yes/no” and the portfolio is equal-weighted across whichever members of the 20 currently pass the screen.

The 20 names: SN2 D-Sperse · SN3 Templar · SN4/51/64 Targon/Llum/Chutes (compute) · SN13 Data Universe · SN17 404-Gen · SN18 Zeus · SN32 ItsAI · SN33 Ready AI · SN34 Bitmind · SN44 Score · SN46 RESI · SN50 Synth · SN56 Gradients · SN62 Ridges · SN68 Metanova · SN75 Hipplus · SN85 Vidalo · SN93 Bitcast.
Why no algorithmic ranking: the whole point of this arm is to ask “does fund-grade curation beat algorithmic top-K?” Adding a scorer on top would contaminate the test.
K = 20 (or however many of the 20 pass the screen) — empirically often 8–12 in any given snapshot, so per-name weight lands around 8–12% (well under the 25% cap).

7. SN0 root staking — the “TAO yield” component

Root staking (netuid 0) is fundamentally different from any subnet alpha. There is no AMM pool — TAO staked to the root validator set earns TAO-denominated yield from network emissions. Two of our sleeves (Lewis, Kidd) anchor a large fraction of capital in SN0 because the people behind those allocations view it as the conservative core position: “own TAO with a productive yield, then take risk on top.”

Position model: SN0 quantity is denominated in TAO. The synthetic price is just the live TAO/USD spot — no pool-impact, no fees, no slippage. A buy of $5,000 at TAO=$240 produces +20.833 TAO; a sell does the inverse.
Yield accrual: on every tick we compound the SN0 position by (1 + APY)^elapsed/year where APY = 8% (a conservative read of recent root validator yield; can be adjusted as a setting).
Mark-to-market: SN0 notional = qty × tao_usd. So the SN0 sub-line of equity moves with TAO spot and grows ~0.022%/day from accrued yield.
Why this matters for the test: a sleeve that is 50% SN0 (Lewis) will track TAO+yield closely on its core; the alpha-or-not question only applies to the 50% that's actively rotated through subnet alphas. Same logic at 30% for Kidd.

8. The Lewis allocation (Mark Jeffrey 10x Challenge)

Lewis Smith / Mark Jeffrey publishes their portfolio openly via the jackson-video-resources/bittensor-investing-agent repo. It's a 7-line conservative allocation built on root staking with measured exposure to a small set of revenue-validating subnets. We replicate it as a fixed-weight sleeve — no screener, no scorer.

Netuid	Name	Weight
SN0	Root staking	50.0%
SN64	Chutes	16.5%
SN62	Ridges AI	11.0%
SN4	Targon	9.5%
SN75	Hippius	7.0%
SN68	Nova	3.5%
SN55	Ko / Precog	2.5%

9. The Kidd allocation (DSV concentrated)

Siam Kidd is CIO of DSV Fund — the world's first TAO-exclusive hedge fund. The 6-line concentrated portfolio reflects a tighter, more aggressive read on which subnets have demonstrated revenue or commercial pull. Less SN0, more concentrated risk per name.

Netuid	Name	Weight
SN0	Root staking	30.0%
SN64	Chutes	25.0%
SN62	Ridges AI	20.0%
SN11	Dippy AI	12.0%
SN19	Nineteen	8.0%
SN75	Hippius	5.0%

How Lewis & Kidd differ from the algo arms. Both bypass the screener and scorer entirely — these are fixed-weight reference portfolios from real public allocators. The 5%-of-equity drift threshold still applies, so they only rebalance when prices move enough to push a name off-target. The point is to give the algorithmic sleeves real human-curated benchmarks to beat.

Research Sleeves v1 — 10 new $10K sleeves added 2026-05-02

The five production sleeves above answer the question "does algorithmic subnet selection beat dumb baselines?". The ten sleeves below ask different questions — are there better TAO alpha strategies we should deploy? Each is a fully-isolated $10K paper sleeve so its equity curve is directly comparable to alphanomics + liquidity. Seven are LIVE on existing data; three are DATA_PENDING — the strategy code and equity tracking are wired, but execution waits on a data dependency.

How these were generated. An adversarial code-and-strategy review of the existing system flagged that backtester numbers were inflated by look-ahead bias, then nominated additional alpha families that fit TAO specifically. Three deep-research agents ran in parallel — one each on TAO perpetual-futures venues, dTAO stake-migration mechanics, and crypto microstructure — to ground each new sleeve in literature and primary sources. Each sleeve has a thesis doc in docs/research-2026-05-02-strategies/ with full mechanics, failure modes, and citations.

The ten new sleeves at a glance

Sleeve	Status	Family	Thesis (one line)
`research_funding_basis`	DATA_PENDING	Carry / basis	Capture TAO perp/spot funding when \|funding_8h\| > 50bps
`research_stake_migration`	LIVE	Cross-section, on-chain	7-day net-flow z-score predicts emission-share repricing over 3–14d
`research_pool_tvl_accel`	LIVE	Cross-section, on-chain	Δ²(pool_tao) generalizes flow_acceleration to depth space
`research_blip_fade`	LIVE	Mean reversion (gated)	Reverse-rank 1w return — only in high realized-vol regimes
`research_xs_momentum`	LIVE	Cross-section momentum	Pure WML30 — A/B baseline for the alphanomics composite
`research_emission_event`	DATA_PENDING	Event-driven	Tilt to subnets with halving in next 72h
`research_tao_btc_ratio`	DATA_PENDING	Pairs / mean-rev	TAO/BTC z-score sets TAO target weight
`research_pool_obi`	LIVE	Microstructure (AMM analog)	(buy_v − sell_v)/total_tao predicts 24h returns
`research_macro_tilt`	LIVE	Regime overlay	TAO 30d trend gates alphanomics composite to 0/0.5/1.0×
`research_regime_ensemble`	LIVE	Meta / regime	RV ratio switches between alphanomics and blip_fade

Reading order: if you want only the headline, scan the table. Each sleeve below has a collapsible block with thesis, mechanics, failure modes, edge expectation, and citations. The same content lives in docs/research-2026-05-02-strategies/<sleeve>.md so you can grep the source.

Sleeve detail

1. research_funding_basis — Carry / basis — DATA_PENDING

Thesis. When TAO perpetual-futures funding rates diverge meaningfully from neutral, a delta-neutral spot/perp position captures the funding without taking directional risk.

Mechanics.

funding_8h ≥ +50 bps → long $5K spot TAO + short $5K TAO perp (collect funding)
funding_8h ≤ −50 bps → short $5K perp + long $5K spot (pay funding, get squeeze)
|funding_8h| ≤ 15 bps → flat (cost-of-capital exceeds expected return)

Why DATA_PENDING. Need a 5-minute funding-rate poller for at least one venue. All eight CEXs that list TAO perps offer public-no-auth APIs (Binance fapi, Bybit v5, OKX, Hyperliquid, dYdX v4, Bitget, Gate, MEXC). Recommended starter: Hyperliquid (1h granularity, free).

Failure modes. Venue-specific liquidation cascades flip funding violently — entry needs a persistent-threshold filter (3 of 4 consecutive obs). Borrow rate on the short leg can negate funding edge; track net yield, not gross. Listing risk if a venue delists TAO.

Edge expectation. Realistic Sharpe target 0.7–1.5 net (vs ~7–12 net for BTC carry per Christin et al.; altcoins discount heavily for liquidity, listing, and borrow).

References. BIS WP 1087 "Crypto Carry" (2023); Christin et al. "The Crypto Carry Trade" (CMU); He, Manela, Ross arXiv:2212.06888; "Funding Rate Arbitrage on CEX and DEX" (ScienceDirect 2025).

Full doc: docs/research-2026-05-02-strategies/01-funding-basis.md

2. research_stake_migration — Cross-section, on-chain — LIVE

Thesis. Post-dTAO (Feb 2025), per-block AMM swaps move alpha-token price instantly while the 30-day-half-life net-flow EMA reprices emission share gradually. The multi-week alpha window between immediate price impact and gradual emission repricing is the trade.

Mechanics. For every subnet, compute z-score of net_flow_7_days cross-sectionally and percentile-rank the absolute level. Composite = 0.6·z_norm + 0.4·pct_level. Filter out subnets where |net_flow_7d| < 0.25% of total_tao (smaller flows absorbed by AMM curvature). Top-K=6 by composite, equal-weight, 25% per-name cap. Rebalance every 30-min tick.

Why this fits Bittensor. Pre-dTAO, emission share was governed by validator votes. Post-dTAO, delegators directly drive emission via AMM swaps and the protocol uses a 30-day-half-life EMA on net flows. That structural lag is guaranteed — the only question is whether 30-min snapshot cadence is fast enough to capture it.

Failure modes. Whale unstake events look like negative migration; mitigated by long-only top-K. Round-tripped flows generate false 1-day signal; the 7-day window helps. Concentration risk if one large validator dominates flow data.

Edge expectation. Lead-lag should manifest over 3–14 days for price. Reasonable target Sharpe vs alphanomics: +0.3 to +0.7 if the structural mechanism delivers; potentially negative if our cadence is too slow. Real read after ≥30 days of forward IC capture.

References. Bittensor Emissions docs; dTAO whitepaper; taostats Stakeholder Emissions Root vs Alpha.

Full doc: docs/research-2026-05-02-strategies/02-stake-migration.md

3. research_pool_tvl_accel — Cross-section, on-chain — LIVE

Thesis. Pools whose total_tao depth is accelerating (Δ² > 0, scale-normalized) outperform on 7-day horizons, capturing both organic stake growth and reflexive feedback (price up → more swaps in → more depth).

Mechanics. Pull 7 days of subnet_metrics.pool_tao per subnet. First diff = inflow rate; second diff = acceleration. Normalize by current pool size for scale-invariance. Top-decile clip to suppress rug-style spikes. Cross-sectional percentile rank → top-K=6.

Why this generalizes. The production alphanomics composite weights flow-space acceleration at 0.05. This sleeve does the same in depth space. A/B vs alphanomics tells us whether depth acceleration adds incremental signal.

Cold-start. Subnets with <3 historical snapshots score neutral. Three days of capture means ~140+ snapshots per subnet — well within the 7-day lookback window. Until the DB has accumulated enough history, the sleeve will rank near-randomly.

Failure modes. Rug-style spikes (one whale dumps 1M TAO into a 100k pool then exits) generate false-positive acceleration; top-decile clip is the partial defense. Pool-genesis bias (brand-new subnets show high normalized acceleration); existing screener catches the worst.

Edge expectation. Should resemble Loop 5 (xs momentum) but with a 12–48h lead. Realistic IC 0.1–0.3 on 7-day forward returns; Sharpe ~0.5–1.0 net.

Full doc: docs/research-2026-05-02-strategies/03-pool-tvl-accel.md

4. research_blip_fade — Mean reversion (RV-gated) — LIVE

Thesis. Cross-sectional reverse-momentum on subnet pool prices clears costs only in high-realized-volatility regimes. Outside that regime, "blip fading" loses to fees + slippage. This is the cost-aware, coded version of "scrape profit on the blips."

Mechanics. Network-level RV = median of per-subnet 7-day log-return stdev. Compute its percentile in the cross-sectional spread. Regime gate: if RV percentile < 0.70, all-cash. Otherwise rank subnets by inverse 1-week return (most-down get highest rank), top-K=6. Filter |1w return| > 50% to avoid failing/pumping subnets.

Cost reality. 60bps × 2 + 30bps slippage = 150bps round-trip. Mean-rev needs 200bps move at 62.5% WR or 400bps at 56% WR to clear. At higher fee tiers (25bps taker), 200bps move at 60% WR breaks even.

Why the regime gate. Caporale & Plastun (2019): short-term reversal in BTC strongest after >2σ moves. Cartea/Jaimungal/Penalva (2015): mean-reversion needs vol expansion to clear bid-ask + fees. Bollerslev/Patton/Quaedvlieg (2016): standard HAR-RV reference.

Failure modes. Late-stage trend continuation looks identical to capitulation blip — RV gate is partial defense. Subnet death (a "blip" because it's failing). Coordinated network-wide dumps.

Edge expectation. 0.8–1.5 Sharpe in vol-expansion regimes; 0.0–0.4 over a full cycle (high-vol regimes are 25–35% of any year).

Full doc: docs/research-2026-05-02-strategies/04-blip-fade.md

5. research_xs_momentum — Cross-section momentum — LIVE

Thesis. Cross-sectional 1-month momentum (WML30) is the simplest factor with proven Bittensor edge per Maymin (2026). Run it solo as a clean A/B vs the alphanomics composite.

Mechanics. Rank universe by price_change_1_month. Top-K=6 equal-weight, 25% per-name cap. No filters beyond pool-screener defaults.

Why this exists alongside alphanomics. Production alphanomics weights momentum_1m at 0.15 alongside 8 other components. This sleeve answers: do the other components add value over momentum alone? If research_xs_momentum matches or beats alphanomics net of costs over 30+ days, the composite is over-engineered.

Failure modes. Classic momentum crashes: when regime flips, highest-momentum names lead the drawdown. Long-only, so worst case is 30–50% drawdown at a regime flip. Mean-reversion regimes hurt momentum; that's why research_regime_ensemble exists.

Edge expectation. Sharpe 0.4–0.9 per Maymin's t-stats translated to net-of-cost. Drawdown 20–35% at worst regime flip. Win rate 52–55%; alpha is in winner magnitude, not rate.

References. Maymin (2026, arXiv:2603.29751); Jegadeesh & Titman (1993, J. Finance).

Full doc: docs/research-2026-05-02-strategies/05-xs-momentum.md

6. research_emission_event — Event-driven — DATA_PENDING

Thesis. Bittensor emission halvings are supply-triggered (not block-triggered). Pre-halving anticipation periods historically produce positive perp basis + alpha pump that fades on "sell the news" — capturable by tilting toward subnets with imminent emission events.

Why DATA_PENDING. The on-chain emission event calendar requires a chain scraper. The data structure exists (EMISSION_EVENT_CALENDAR: dict[int, list[int]] in the strategy module) but is not populated. Once a scraper extracts upcoming halving thresholds per subnet, the sleeve activates automatically.

Mechanics (when live). For each subnet, compute next emission-event ts (extrapolated from current emission rate vs cumulative-supply threshold). Include if 0 < hours_until ≤ 72; weight ∝ 1 / hours_until. Top-K=6, hold through event + 24h, then exit.

Failure modes. Anticipated events get priced in earlier as more participants run the same strategy — 72h window may shrink. Whales know the schedule too and front-run our window. Calendar slippage as emission rates change.

Edge expectation. Modest. Realistic 0.3–0.6 Sharpe during active halving periods, near-zero in cooldown periods.

Full doc: docs/research-2026-05-02-strategies/06-emission-event.md

7. research_tao_btc_ratio — Pairs / mean-rev — DATA_PENDING

Thesis. TAO has historically traded with high beta (~1.4) to BTC. The TAO/BTC ratio mean-reverts on a multi-week horizon, so a TAO position that is neutral on BTC captures TAO-specific alpha (subnet emission cycle, validator dynamics) without taking the broader crypto-cycle bet.

Why DATA_PENDING. Need BTC/USD price history (CoinGecko free tier supports it; not currently fetched) and a single-asset paper engine (or fix to the disabled tao sleeve cash accounting).

Mechanics (when live). Compute ratio_t = (TAO/USD)/(BTC/USD); z = (ratio − mean_60d) / std_60d. Map to TAO target weight: clamp01((−z + 1.5) / 3). z = −1.5 → 100% TAO; z = 0 → 50%; z = +1.5 → 0%. Single-asset on TAO/USD spot. Daily rebalance.

Why this matters. Existing subnet sleeves all track TAO-beta strongly because subnet alpha is priced in TAO. During a BTC drawdown that drags TAO down, every subnet sleeve loses TAO-beta even if subnet scoring is right. Ratio sleeve is the cleanest way to isolate TAO-specific alpha.

Failure modes. Beta drift (TAO/BTC beta is not constant). Secular ratio decay (TAO has structural issuance; BTC halves every 4 years). Liquidity asymmetry — don't size beyond ~$10K notional.

Edge expectation. Low conviction. 0.4–0.8 Sharpe in cyclical years, near-zero in trending years. Best property: decorrelated diversifier vs all subnet sleeves.

References. Gatev, Goetzmann & Rouwenhorst (2006, RFS) on pairs trading; Liu & Tsyvinski (2018) on cryptocurrency beta.

Full doc: docs/research-2026-05-02-strategies/07-tao-btc-ratio.md

8. research_pool_obi — Microstructure (AMM analog) — LIVE

Thesis. Subnet AMM pools have no L2 order book, but (buy_v − sell_v)/total_tao with depth-normalized pressures is a faithful multi-level OBI analog and predicts contemporaneous + 24h-leading subnet returns.

Mechanics. buy_pressure = tao_buy_volume_24_hr / total_tao; sell_pressure = tao_sell_volume_24_hr / total_tao; pool_obi = (buy_pressure − sell_pressure) / (buy_pressure + sell_pressure). Cross-sectional rank, top-K=6.

Why volume-over-depth. Raw buy/sell ratio (what alphanomics' buy_sell_pressure uses) defaults to 0.5 in inactive subnets and ignores pool size. Depth normalization means a $100k buy into a $50k pool counts more than a $100k buy into a $10M pool.

What microstructure research shows. Multi-level OBI: 60–80% R² on contemporaneous mid-price moves in equities (Cont/Kukanov/Stoikov 2014); 0.3–0.5 R² in crypto replications (Kolm et al. 2023). Coinbase L3 not publicly available — we cannot detect spoofing via order lifetime, but AMM swaps are deterministic and on-chain (no spoofing).

Failure modes. 30-min latency drag (some alpha already priced in by the time we trade). Whale dump-and-pump cycles distort OBI; depth normalization actually amplifies this. MEV/sandwich risk in live mode (mitigated via private RPC).

Edge expectation. Should outperform raw buy/sell pressure modestly. 0.3–0.7 Sharpe as standalone. Best paired with xs_momentum — OBI is faster than momentum, could front-run entries.

References. Cont/Kukanov/Stoikov (2014); Silantyev (2019); Cartea/Drissi/Monga (2023) on AMM microstructure.

Full doc: docs/research-2026-05-02-strategies/08-pool-obi.md

9. research_macro_tilt — Regime overlay — LIVE

Thesis. Subnet alpha portfolios are dominated by TAO-beta during TAO bear markets. Gating gross exposure by the TAO 30-day trend isolates the alpha-vs-TAO performance and protects sleeve PnL during macro drawdowns.

Mechanics. Pull TAO/USD 30-day return from prices. Trend gate: 30d return > 0 → 1.0× (full alphanomics composite, $10K notional); ∈ [−10%, 0] → 0.5× ($5K + $5K cash); < −10% → 0× (all cash). Run production subnet_scorer.score() at gated equity.

Why TAO-trend (not stablecoin-supply). Stablecoin-supply growth is the long-term direction (USDC + USDT inflows correlate with risk-on regimes), but free APIs are operationally heavy. TAO-trend is a cheap, faithful proxy: rising TAO = risk-on for subnet alpha.

Failure modes. Whipsaw at the 0% boundary in sideways regimes (planned mitigation: hysteresis — exit at −2%, re-enter at 0%). Late entry/exit (30-day trend confirms regimes after they're well underway). TAO-vs-network divergence (subnet-specific events the gate doesn't see).

Edge expectation. Lower headline Sharpe than alphanomics in trending years; higher Sharpe in cyclical years. Should reduce max drawdown by ~30–50% if the gate fires correctly during major bear regimes. 0.5–1.0 Sharpe vs alphanomics depending on year regime mix.

References. Faber (2007) "A Quantitative Approach to Tactical Asset Allocation" — simplest moving-average regime gate.

Full doc: docs/research-2026-05-02-strategies/09-macro-tilt.md

10. research_regime_ensemble — Meta / regime — LIVE

Thesis. Switch between alphanomics composite (calm regime) and blip-fade (vol-expansion regime) based on TAO realized-volatility ratio. Meta-strategy whose alpha comes from correctly timing the switch.

Mechanics. Pull TAO log returns from a 90-day window. Compute current RV (recent ~45d) and historical RV (older ~45d). rv_ratio = current_rv / historical_rv. Map to regime:

rv_ratio > 1.5 → HIGH → 100% blip_fade
rv_ratio < 0.5 → LOW → 100% alphanomics
else → MID → 50% alphanomics + 50% blip_fade (combined targets)

The most uncertain sleeve. Its alpha depends on the regime detector being right. After 60+ days, three comparisons matter:

If regime_ensemble ≥ both xs_momentum and blip_fade standalone → regime detector adds value.
If it tracks a static 50/50 of alphanomics + blip_fade → detector adds nothing.
If it underperforms a static 50/50 → detector is destroying value (mistiming switches).

Failure modes. Regime detector lag (RV percentile crosses thresholds after the regime has changed). Whipsaw at boundaries (planned: hysteresis — 1.6 to enter HIGH, 1.4 to exit). Mid-regime weighting (50/50 is a guess; continuous blend planned).

Edge expectation. If detection works: 0.8–1.4 Sharpe combined. If neutral: 0.4–0.8 (matches static blend). If wrong: <0.3 (whipsaw losses).

References. Ang & Bekaert (2002); Kritzman, Page & Turkington (2012); López de Prado (2018) on purged CV for regime-aware backtests.

Full doc: docs/research-2026-05-02-strategies/10-regime-ensemble.md

How these sleeves get evaluated

Adversarial review found a CRITICAL look-ahead-bias bug in the existing backtester (engine.py:61-66, subnet_engine.py) that inflates every reported Sharpe. Until that's fixed, no scripts/backtest*.py number is informative. So these sleeves are evaluated only on forward IC — real paper-trading from t=0. After ≥30 days of paper data, scripts/ic_analysis.py can compute IC per sleeve and per signal component.

Day 7+: first signal-direction read (still noisy). Look for sleeves drifting wildly off the alphanomics baseline — early evidence of differential edge.
Day 30+: first credible Sharpe estimate. Sleeves with Sharpe < 0.3 should be flagged for removal review.
Day 60+: regime ensemble vs static-50/50 comparison resolves whether regime detection is value-added.
Day 90+: graduate top-3 to a candidate live-capital list — only after the look-ahead-bias bug is also fixed.

Total new paper capital deployed: $70,000 ($10K × 7 LIVE sleeves). The 3 DATA_PENDING sleeves track a flat $10K equity until their data dependency lands; activating each is roughly 1–5 days of engineering. None of these are real money.

Inception-to-date P&L by sleeve

15 paper sleeves: 5 production + 10 research v1. Each started with $10,000. Numbers below are live from the SQLite DB at report build time.

Equity curves

Legend: dashed segments = synthetic backfill (strategy replayed against historical snapshots since 2026-04-19, look-ahead-safe). Solid segments = live paper trading. Backfill is not a live result — treat as a directional fairness check, not a Sharpe number.

Per-sleeve detail

Click a sleeve to expand current open positions and the last 8 trades. Filtered to match the chips above.

What we’re actually trying to learn — and how

Paper trading is cheap. The point is not “simulate making money,” it’s to generate the dataset that lets us decide whether to risk real money — and if so, on which signals.

The hypotheses being tested

#	Hypothesis	Falsified if…
H1	SMB is real on dTAO subnets in the current regime	size_inverse component has near-zero forward IC after 30 days of capture
H2	Net stake flow predicts forward returns (the Taoflow story)	alphanomics_short / _med components have IC ≈ 0 or negative
H3	Composite scoring beats raw liquidity baseline	alphanomics sleeve underperforms liquidity sleeve over a full month
H4	Subnet selection beats just owning TAO	all five sleeves underperform spot TAO total return (note Lewis/Kidd already include partial TAO exposure via SN0)
H5	Curated fund picks (DSV broad) beat algorithmic top-K (alphanomics)	DSV-20 sleeve underperforms alphanomics over a full month — would suggest the algorithm captures whatever edge the fund picks have, with no expert premium
H6	Concentrated discretionary (Kidd) beats both broad-curated (DSV-20) and conservative-discretionary (Lewis)	Kidd is dominated by either over a full month — would suggest concentration adds risk without return
H7	Active subnet selection adds value over the SN0 yield baseline	any sleeve underperforms a hypothetical 100%-SN0 position (TAO total return + 8% yield) over a full month
Research v1 sleeves (added 2026-05-02)
H8	Funding-rate basis is a real edge on TAO perps	`research_funding_basis` realized Sharpe < 0.3 net of fees + borrow over 90 days once data is wired (vs ~0.7–1.5 target)
H9	Stake-migration flow leads emission repricing in 3–14d	`research_stake_migration` tracks alphanomics ±20bps daily over 30 days — i.e. no incremental signal from the flow lag
H10	Pool-TVL acceleration (depth space) beats flow-acceleration alone	`research_pool_tvl_accel` underperforms `subnet_alphanomics` by >30% over 30 days (depth signal adds nothing over the existing flow_acceleration component)
H11	"Scrape the blip" only works in high-vol regimes	`research_blip_fade` deploys for <15% of ticks over 60 days (regime gate too strict) or deploys often but loses to fees by >5% (gate too loose)
H12	The alphanomics composite adds value over pure WML30 momentum	`research_xs_momentum` matches or beats `subnet_alphanomics` within ±50bps daily over 30 days — would mean the 8 non-momentum components are noise
H13	Pre-halving anticipation produces capturable basis	`research_emission_event` sleeve realized Sharpe < 0.3 over 90 days once data is wired (with at least 3 emission events in window)
H14	TAO/BTC ratio is a useful diversifier from subnet sleeves	`research_tao_btc_ratio` daily-PnL correlation to `subnet_alphanomics` > 0.7 over 60 days once wired (no diversification benefit)
H15	Pool-OBI predicts 24h returns	`research_pool_obi` realized IC < 0.05 on 24h forward returns after 30 days (signal too noisy or too slow at our cadence)
H16	TAO-trend gating reduces drawdown without giving up too much upside	`research_macro_tilt` max drawdown over 60 days is within 5pp of `subnet_alphanomics` — gate isn't actually protecting or total return is >30% below alphanomics (gate too pessimistic)
H17	Regime-switching is value-additive over a static blend of components	`research_regime_ensemble` underperforms a static 50/50 of `subnet_alphanomics` + `research_blip_fade` over 60 days — regime detector is destroying value via mistimed switches

How we measure it

Test windows

Days 1–7: shake-out. Watch for execution bugs, weird snapshot data, edge cases. No conclusions drawn.
Days 7–14: first IC numbers. Run scripts/ic_analysis.py to compute Spearman correlation between each scoring component and forward returns. Identify components with |IC| > 0.05 and consistent sign.
Days 14–30: reweight. Drop or down-weight components with no predictive power. Add Taoflow EMA proper (we have enough history at this point).
Day 30+: regime test. Compare sleeves head-to-head over a full month containing varied TAO macro conditions. Decide: graduate to Phase 2, iterate, or kill.

Why the directive is “more trades, not fewer”

Real-money strategies want to minimize trades to save fees. This system wants to maximize trades to maximize the labeled dataset — every trade is one row of (signal at t, return by t+1). The 5%-of-equity threshold is the natural friction; we don’t artificially gate further.

Risks we’re explicitly tracking

Survivorship bias. Deregistered subnets disappear from the API. We need a separate snapshot table for failed subnets — currently a TODO.
Capacity. All findings are valid only at $10K AUM scale. Maymin is explicit that the SMB factor decays sharply with size.
Regime risk. The first 30 days might land in a strong bull or strong bear for TAO. The 30-day backtest already showed all strategies losing money in the post-Covenant window. We need multi-regime data before any go/no-go.
Free-tier API. 5 req/min throttle means a tick takes ~25s. Acceptable today; if we add new pull endpoints, recount the budget.
Single-writer SQLite. Fine at this scale; replace if we ever fan out the runner.

Bottom line. The bot is not a money machine — it’s an instrumented experiment whose output is a clean answer to “which subnet signals actually work, and at what scale?” Real-money allocation is a downstream decision that lives or dies by the data this dashboard accumulates.

Generated by tao-quant · data window from inception to report build time · auto-refresh by re-running the build script.