Scoring Reference

How The Bot Decides

Every 30 minutes the rebalancer walks through the same pipeline for every Bittensor subnet:

Pre-filters — hard rules that throw out anything obviously unsafe (too expensive, too thin, crashing, etc).
Scoring — survivors get a 0–100 score across 7 components. Higher = stronger trend + stronger buy pressure.
Tier mapping — score becomes a tier (Conviction / T1 / T2 / T3). Tier sets the target size.
Tier caps — overlay rules that can downgrade or skip an otherwise-good score (e.g. "too far above EMA → skip").
Entry filters — last set of idiot-checks before any buy fires (don't buy parabolic, don't buy below EMA, etc).
Ladder pyramid — first buy is capped at T2 size. Position scales up one tier per 4h candle only if the trend keeps confirming.
Trade — buys delegate to the highest-yield validator on that subnet.

Once a position is open, monitor.py runs every 15 minutes and checks exit rules (trailing stop, score decay, flow reversal, EMA breakdown, take-profit checkpoints, hard stop loss, kill switch).

    Total Score = Momentum (40 pts max) + Buy Pressure (60 pts max).

    Momentum measures price trend strength. Buy Pressure measures capital inflow. Buy Pressure is weighted more heavily because flow leads price.

Score Components (out of 100)

Component	Category	Max Pts	What It Measures	How It Works
Slope Acceleration	Momentum	11	Is the price trend accelerating?	Compares recent 4h candle slope to earlier slope. A steepening uptrend scores high. Flat or decelerating trends score low. Uses a 325× multiplier on the slope difference.
Price Position 7D	Momentum	11	Where is the price within its 7-day range?	`(current − 7d_low) / (7d_high − 7d_low)`. Near the top of the 7-day range = strong. Near the bottom = weak. Scores linearly.
Monthly/Weekly Momentum	Momentum	10	Is the trend sustained across timeframes?	Combines 30-day and 7-day price changes. Rewards subnets where BOTH timeframes are positive (sustained trend, not a single spike). Penalises divergence.
EMA Freshness	Momentum	8	How recently did price cross above the EMA?	Uses the 18-period 4h EMA (≈72h trend). Price above EMA = bullish. Recently crossed above = extra points (fresh breakout). Below EMA = 0–1 points.
Flow Ratio	Buy Pressure	34	Are buyers outweighing sellers?	24h TAO buy volume vs sell volume. Strong buy-to-sell ratio (e.g. 2:1) scores high. Also factors wallet counts and volume consistency. This is the single heaviest-weighted component — capital flow is the strongest leading signal we have.
Flow Acceleration	Buy Pressure	16	Is buying pressure accelerating?	Compares recent flow to earlier flow. If buying is ramping UP (not just steady), it scores high. Uses a 600× multiplier on the acceleration. Catches early-stage buy surges before they show in price.
Wallet Ratio	Buy Pressure	10	More unique buyers than sellers?	Distinct buyer wallets vs seller wallets in 24h. A high buyer/seller ratio means broad-based demand (not just one whale). Capped at 10 points.

Component values use the current 4h candle in progress (use_rolling_current_candle: true), so the score reflects what's happening live, not just the last closed candle.

Tier Allocation

Score maps to a tier. Tier sets the target percentage of the portfolio that position should occupy. Anything scoring below 30 is dropped — no allocation.

Tier	Score Range	Target %	Per-Position Hard Cap	Entry Behaviour
Conviction	75 – 100	7%	10%	Forced entry: can evict the lowest-conviction open position to make room, unless the evict target is protected (active trailing stop or > 30% ROI).
Tier 1	50 – 74	5%	10%	Standard entry when a slot is available.
Tier 2	40 – 49	3%	10%	Standard entry when a slot is available.
Tier 3	30 – 39	1.5%	10%	Standard entry when a slot is available. Tier 3 is "probe size" — a small toehold while we wait for the trend to confirm.

    Static tier sizing (pro-rata OFF). Each position gets exactly its tier target — we do not scale up to fill capital. If top scorers don't fill the book, the remainder sits as cash rather than inflating lower-conviction picks. Max open positions: 15. Gas reserve: 0.20 TAO always held back.
  

    Position cap (safety). If a single position ever grows past 20% of the portfolio (e.g. a runner), it gets trimmed back to 10%. Prevents one winner from becoming the whole book.
  

Ladder Pyramid ladder_pyramid_v1 · enabled

Sizing rule layered on top of the tier table above. Borrowed from Gordy: enter small, scale up only as the trend keeps confirming.

First buy is capped at Tier 2 (3%) — even a Conviction-grade score gets a T2-sized first entry. We're proving the trend with real skin, not betting the farm on one snapshot.
One step per 4h candle close. At every 4h close (00/04/08/12/16/20 UTC), if the score still says ≥ next tier, the position upgrades one step: T2 → T1 → Conviction. Never two steps in one candle.
Conviction is the ceiling. Never sized larger than 7%.
No auto-downgrade. If the score drops, exits own the shrink (trailing stop, score decay, flow reversal). The ladder only goes up.
TP-lock honored. If we've already taken profit at a tier on this position, the ladder won't re-grow back to that same tier — only above it.

Why: scoring decides if trending. Sizing decides conviction level. Confirmation comes from time + continuation, not a single snapshot.

Pre-Filters (Hard Rules Before Scoring)

A subnet must pass all of these to be eligible for scoring. Anything that fails is dropped — no further evaluation.

Filter	Rule	Why
Price Cap	Price ≤ 0.04 TAO	Avoid mature/expensive subnets where 5% moves take days.
Price Floor	Price > 0	Must have a real, non-zero price.
Min Pool Depth	Pool ≥ 2,800 TAO	Enough liquidity to enter and exit without crushing the price.
Max Pool Depth	Pool ≤ 50,000 TAO	Too large = too mature, moves are slow.
Gini Coefficient	Gini ≤ 0.85	Not too concentrated in a few wallets (whale-controlled).
Monthly Pump Cap	30d change ≤ 500%	Already pumped — late to the party.
All-Zero Guard	Must show non-zero price changes	Skip dormant/dead subnets.
Accelerating Sell	NOT (1d < −5% AND 7d < −5%)	Don't catch a falling knife that's accelerating down.
Flat Momentum	NOT (30d < 3% AND 7d < 3%)	Going nowhere — no edge.
Dual Downtrend	NOT (7d < 0 AND 30d < 0)	Confirmed downtrend on both timeframes.
Structural Decline	30d change > −10%	In structural collapse.
Day Crash	1d change > −20%	Flash-crashing — wait for it to find a floor.
Immunity Period	Not in startup/immunity mode	Brand-new subnets are protected from emission and not yet tradeable normally.
Blacklist	Not in `config/blacklist.json`	Permanent skips (broken subnets, known scams, etc).

Entry Filters (Pre-Buy Idiot Checks)

Even after a subnet scores well, these filters block the actual buy if the entry timing looks unsafe. All run on the surviving candidates after scoring.

Filter	What It Checks	Why
EMA Direction	Price must be above the 18-period 4h EMA.	Don't buy into a downtrend. The 18 EMA on 4h candles ≈ a 3-day trend filter.
EMA Spike	Price must have been above the EMA for at least 2 consecutive 4h closes.	One candle above isn't a trend — it's a wick. Wait for confirmation. Bypassed by dip-buy below.
Parabolic Guard	Latest 4h candle range can't exceed 1.5× the recent average range.	A single super-wide candle is the top of a parabolic move more often than the start of one.
RSI Overbought	RSI(25) must be below 75 — relaxed to 82 for mature trends (12+ candles above EMA).	Don't buy at peak exhaustion. Mature trends earn more rope.
Price Drift	Live price can't have drifted more than 5% from the price used in scoring.	If price moved a lot between scoring and the buy, our data is stale — abort and re-evaluate next cycle.
Re-entry Cooldown	Must wait 4 hours after selling a subnet before re-buying it.	Prevents whipsawing into the same name on noise.
BB Squeeze	If Bollinger Bands are squeezed (< 5% width), require a clear breakout above the upper band before entry.	Avoid buying inside a coil — wait for direction.

Tier Caps (Score Overlays)

Tier caps run after scoring but before the buy. They can either downgrade a position's tier or skip the entry entirely, even when the score is good. The strictest applicable cap wins.

Cap	Status	Rule	Effect
Extended Above EMA extended_above_ema_block_v1	enabled	Price more than 7% above the 10-period 4h EMA.	Hard skip. Catches "we're chasing a candle that already ran" entries. Sweep showed it blocks 11 of 99 entries, removing 3 of the 12 worst KDJ-only losers.
BB Top Block bb_top_block_v1	disabled	Price above the upper Bollinger band.	Would cap entry size to T3 (small probe) rather than hard-skip. Kept on the shelf in case we want softer behavior.
Bear Structure Skip bear_structure_skip_v1	disabled	10 EMA below 18 EMA AND last bullish crossover more than 30 candles ago.	Would hard-skip "dead structure" subnets. Off pending more data — risk of skipping early bottom reversals.

Dip-Buy Bypass dip_buy_v1 · enabled

Lets us enter on a fresh "crossing up through the EMA from below" setup that the standard EMA Spike filter would otherwise block. The earliest, lowest-risk entry pattern Gordy uses.

Triggers a bypass when ALL of these hold:

10-period EMA is rising (slope ≥ 0 over the last 6 candles).
Current close is above the 10 EMA but within 1.5% of it (price is hugging the EMA from above, not extended).
One of the last 3 closes was below the 10 EMA — confirms it's a fresh cross-up, not a stale uptrend.
Current candle is green (close > open).
Flow ratio ≥ 1.2 (buyers actually showing up).

When all five fire, the EMA Spike filter is skipped for that subnet on that candle. EMA Direction, Parabolic Guard, RSI, drift, cooldown, and tier caps all still apply.

On the dashboard, blocked rows that would have qualified for the dip-buy bypass show a cyan dip-elig tag, so you can see what we'd have caught if another filter hadn't fired.

Trend Confirmed trend_confirmed_v1 · live

Layered scoring + exit logic that tracks whether a position is riding a clean, confirmed trend off either the 10 EMA or the 18 EMA. Lets us be patient with real trend riders and aggressive about cutting fake ones.

Primary EMA

For each subnet we pick a primary EMA (10 or 18) based on which one price has been bouncing off cleanly. "Bouncing" = a candle body within 0.8% of the EMA, followed by a green candle that leaves it. Wicks don't count.

Pre-Trade Trend Strength

Before any buy, we look at the last 30 4h candles and check whether the setup qualifies as a confirmed trend on entry:

Path A (immediate confirmed): trend ≥ 18 candles old, ≥ 2 bounces off the primary EMA, zero 4h closes below the 18 EMA. Position enters in CONFIRMED state.
Path B (post-entry): doesn't qualify at entry but earns CONFIRMED later if ≥ 2 days, ≥ 2 bounces, zero closes below 18 EMA accumulate while we hold.

Score Bonuses

+10 pts — bounce off the primary EMA.
+5 pts — bounce off the secondary EMA.
+6 pts — holding cleanly above the primary EMA.
−10 pts — closed below the 18 EMA (trend integrity broken).

Parabolic Entry Filter

If price is more than 10% above the 10 EMA, entry is blocked — we're not chasing a parabolic move regardless of score.

Three-Zone Flow Reversal Exit

Replaces the legacy flat flow-reversal exit. Behavior depends on where price sits relative to the primary EMA:

Upside zone (> 8% above primary EMA): 3 consecutive outflow candles + score below the tier threshold → exit. Captures profit-taking when a parabolic move starts to fade.
Middle zone (between the EMA and +8%): no flow-reversal exit. The position rides — either confirms upside or breaks downside.
Downside zone (latest 4h close at or below the 10 EMA after a prior close above it): three-candle window around the breach all outflow AND score ≤ 30 → exit. Catches positions that bleed below the EMA while inflow blips reset the legacy counter.

Hard Confirmed Exit

If the position is CONFIRMED and a 4h candle closes below the 18 EMA, the position is full-sold immediately. Trend is broken.

Exit Rules

Monitor cycles every 15 minutes. A position is held for at least 2 hours minimum before any exit rule (other than the hard stop loss / kill switch) can fire. Then the rules below evaluate in order.

Trailing Stop (tier-dependent)

Activates once a position has hit +10% ROI. Then sells if price drops by the trail amount from the running peak:

Tier at activation	Trail %
Conviction / Tier 1	14%
Tier 2	10%
Tier 3	7%

Progressive Tightening

As gains grow, the trailing stop tightens automatically — we protect a bigger share of the move:

ROI	Trail tightens to
+55%	12%
+75%	8%
+90%	6%

Take-Profit Checkpoints (auto-trim)

ROI Hit	Trim	Notes
+14%	15% of position	First skim — locks in some win even if it reverses.
+29%	15% of position	Second skim.
+60%	20% of position	Big skim — by here we've locked 50% of the position. The rest rides on the trailing stop.

After a TP trim, the rebalancer cannot top the position back up to the same tier — only to a higher tier earned via ladder upgrade. Stops trims from being silently undone.

Score Decay (tier-dependent)

If the score drops below the threshold for 2 consecutive checks, the position is sold — its fundamentals have deteriorated.

Tier at entry	Decay threshold
Conviction / Tier 1 / Tier 2	20.5
Tier 3	22

Flow Reversal (tier-dependent)

3 consecutive 4h candles of net outflow AND score has slipped below the tier threshold → exit. (See Trend Confirmed above for the zone-based variant that's live now.)

Tier	Score must be below
Conviction / Tier 1	28
Tier 2	30
Tier 3	35

EMA Breakdown

Price closes below the 18-period 4h EMA for 5 consecutive candles → exit. The trend is broken — get out.

Hard Stop Loss

Position drops −12% from entry → exit immediately. No min-hold delay applies here.

Position Cap

If a single position grows past 20% of the portfolio (a runner), trim back to 10%. Prevents concentration risk.

Kill Switch

If the portfolio drops −20% in a single day, ALL trading halts until manually reset. Hard safety net.

Validator Delegation validator_delegation_v1 · enabled

Every buy now delegates the stake to the highest-yield validator on that subnet, rather than staying on our own hotkey. That picks up emission yield Gordy was earning and we weren't.

Top validator is chosen per subnet using api.get_best_validator(netuid).
If the API returns no qualified validator, we fall back to delegating on our own trading hotkey (legacy behavior).
Top-ups go to the same validator the position is already delegated to — never split a position across hotkeys, otherwise exits get messy.
Sells call --include-hotkeys validator_ss58 so externally-delegated stakes still unstake cleanly.

Glossary

EMA — Exponential Moving Average. We use the 10-period and 18-period on 4-hour candles. The 18 EMA on 4h candles spans roughly 3 days, so it's the medium-term trend line.
Bounce — a 4h candle whose body comes within 0.8% of an EMA (wick can poke through, body doesn't), followed by a green candle that leaves the EMA. Counts as evidence the EMA is acting as support.
4h candle close — every 4 hours at 00, 04, 08, 12, 16, 20 UTC. Ladder upgrades and score-bonus reassessments are anchored to these.
ROI — return on investment for an individual position, measured against the average entry price.
BP / Flow — Buy Pressure / capital flow. The TAO-volume side of scoring (flow ratio + flow acceleration + wallet ratio = 60 of the 100 points).
Tier 4 (T4) — terminology you may see in chat for "below T3" probes. Currently not a sized tier in this bot — anything scoring below 30 is dropped, not held at T4 size.
Conviction — top tier (score ≥ 75). Only tier allowed to evict an existing position to make room.
Gordy — Siam's reference Gordon bot wallet. We display it side-by-side as a benchmark.

Scoring Model Reference