A long/short strategy: portfolio mathematics and a systematic approach

1. Version and notation

Checked against source and settings on September 14, 2026. Plugin rs_rotation_v1, model original_cov_720h_168h_top5_hourly_v2, policy continuous_target_v2. The URL is version independent; the old address redirects with HTTP 301. This documents checked code, not proof that every setting is currently loaded by the lead account or reproduced by followers.

E is current account equity; f allocation fraction; A allocated capital; G total absolute target notional across both sides; A₀ initial allocated capital. q is contracts and c is base-asset quantity per contract. BTC is benchmark only.

A = fE; G = kA
f = 1; k = 10; L_i = 10 for AVNT; 20 otherwise
A₀ persists; it is not a high-water mark

2. Full universe and data

There are 58 coins: 50 core assets and 8 optional additions — ETHFI, RIVER, AVNT, 币安人生, JUP, 牛来, PONS, MARSCOIN. BTC is a separate benchmark and is not one of the 58. An optional asset needs complete hourly history, at least 50,000 USDT of 24-hour turnover and a spread no wider than 0.5%. Missing history or failed filters exclude it from current selection; a core-asset or BTC data error blocks calculation. If an optional coin is already held, missing history blocks the new plan. Optional coins also require catalog availability and aligned leverage; liquidity is checked again before entry. All 58 coins share one portfolio and risk budget.

History contains 721 closed hourly candles for each included asset and BTC: H = 721×(n+1)×6, with 50≤n≤58. With all 58 assets: 721×59×6. Fields are opening timestamp, open, high, low, close, volume. Candles must be finite, consistent and gap-free; OHLC>0, volume≥0. Missing candles are never filled. Order is core assets, available optional assets, BTC last.

ETH SOL XRP DOGE BNB ADA SUI HYPE PEPE ZEC
LTC FIL WLD TRX XAUT LINK ENA BCH ARB NEAR
TRUMP DOT DASH SHIB WIF ASTER GRAM UNI AVAX ONDO
ORDI ETC OP TAO XLM ATOM XMR FLOKI BEAT FARTCOIN
APT PUMP ICP TIA PENGU LIT PEOPLE AAVE BONK SUSHI
ETHFI RIVER AVNT 币安人生 JUP 牛来 PONS MARSCOIN

D mod 3600 = 0
timestamps = D−721h, …, D−1h
high ≥ max(open,low,close); low ≤ min(open,high,close)
input_hash = SHA256(H: little-endian float64, C order)
spread = (ask-bid)/((ask+bid)/2) ≤ 0.005
optional turnover_24h ≥ 50000 USDT
ticker age ∈ [−5, 30] seconds

3. Returns, beta and residual risk

P is close and B is BTC. Risk uses 720 simple hourly returns. Sample estimates use ddof=1 (denominator 719); beta is centered. BTC variance<10⁻¹² blocks calculation. Residual e has no separate intercept/alpha; σ is its hourly standard deviation, not annualized.

rₜ,ᵢ = Pₜ,ᵢ/Pₜ₋₁,ᵢ − 1, t=1,…,720
r̄ᵢ = Σₜrₜ,ᵢ/720
vB = Σₜ(rₜ,B−r̄B)²/719
βᵢ = [Σₜ(rₜ,ᵢ−r̄ᵢ)(rₜ,B−r̄B)/719]/vB
eₜ,ᵢ = rₜ,ᵢ−βᵢrₜ,B
σᵢ = sqrt[Σₜ(eₜ,ᵢ−ēᵢ)²/719]

4. Score and eligibility

The signal uses 168-hour log returns while beta/σ use 720 simple hourly returns. This explicit mixture is not a log-return regression. Score is not an expected return or success probability.

Eligibility: β∈[0.2,3], σ strictly >10⁻⁶ and volume>0 in at least 90% of the last 720 candles. Valid ineligible assets remain in input. Funding and depth are not score inputs.

Lᵢ = ln(P_last,ᵢ/P_last−168,ᵢ)
scoreᵢ = (Lᵢ−βᵢLB)/max(σᵢ,10⁻⁶)/sqrt(168)
eligible = finite(β) AND 0.2≤β≤3 AND σ>10⁻⁶
           AND count(volume>0)/720 ≥ 0.90

5. Selection and ties

Fewer than ten eligible produces the zero target insufficient_eligible, which may close positions. Data/solver errors do not mean flat: no new target is published.

Stable ascending score sort: short takes first five, long takes last five reversed. Ties use universe order with the long tail reversed. Score sign is not a filter. Zero QP weight does not remove raw membership.

rank = stable_argsort(score[eligible])
short = rank[:5]; long = reverse(rank[−5:])
selected = long + short
s = (+1,+1,+1,+1,+1,−1,−1,−1,−1,−1)

6. Covariance and optimization

S uses raw simple asset returns, not residual e. The matrix mixes 75% S and 25% diagonal. x is absolute G weight, w=s⊙x. Score selects candidates but is absent from the objective: no momentum reward, turnover penalty, funding forecast or expected-profit optimization.

Zero weights are allowed; long/short dollar notionals need not match. No equal-weight or simple-volatility fallback.

Sᵢⱼ = Σₜ(rₜ,ᵢ−r̄ᵢ)(rₜ,ⱼ−r̄ⱼ)/719
Σ = 0.75S + 0.25diag(diag(S))
Q = Σ_selected ⊙ (ssᵀ)
minimize 0.5xᵀQx
subject to Σxᵢ=1; Σsᵢβᵢxᵢ=0; 0≤xᵢ≤0.15
w=s⊙x

7. Numerical solver protocol

Exactly 10 variables, 10×10 matrix, 5+/5− signs, finite values and β∈[0.2,3]. Symmetry allclose uses atol=10⁻¹² and default rtol=10⁻⁵; min eigenvalue≥−10⁻¹². Q scaling and 10⁻¹⁰I are numerical regularization. Initialization above cap+10⁻¹² blocks even if another feasible point exists.

After solver, allocate rechecks finite, |gross−1|≤10⁻⁸, |w·β|≤10⁻⁸, cap≤0.15+10⁻⁹ and wᵢsᵢ≥−10⁻¹⁰. Errors/nonconvergence do not yield a simplified portfolio. The entire preset must match model parameters.

Q̃ = Q/max(max(diag(Q)),10⁻¹²) + 10⁻¹⁰I
bL=mean(β_long); bS=mean(β_short)
x_long=bS/[5(bL+bS)]; x_short=bL/[5(bL+bS)]
C=[1ᵀ; (s⊙β)ᵀ]; Cx=(1,0)ᵀ
repeat at most 500 iterations:
  F = non-active coordinates; require |F|≥2
  g=Q̃x
  solve [Q̃_FF C_Fᵀ; C_F 0][d_F;λ]=[−g_F;0;0]
    using least squares, rcond=10⁻¹²; d_active=0
  if max|d|<10⁻¹⁰:
    z=g+Cᵀλ
    violation=−z_j at lower bound, +z_j at upper
    if no active bounds or max violation<10⁻⁹:
      clip x to [0,0.15]
      require max|Cx−(1,0)|≤10⁻⁸, max(x)≤0.15+10⁻¹⁰
      return x
    else release most violated bound
  else:
    step starts at 1, reduced to first bound along d
    |d_j|≤10⁻¹² does not set a bound
    x←x+step*d; activate bound reached
otherwise: OptimizationBlocked

8. Membership confirmation

Only D−3600 is adjacent. Per side, retain eligible incumbents present in current or previous raw, then add new candidates present in both. Incumbents first, maximum five. First absence removes entry support; two adjacent absences remove accepted membership. Lost eligibility removes immediately.

After a time gap, eligible incumbents remain and new confirmations restart. Overlap retains the previous side (long searched first). Entry requires supported, nonzero correctly signed raw weight and a flat opposite side. Missed hours are not traded retroactively.

adjacent = previous.D==D−3600
raw_prev = previous.raw if adjacent else empty
kept = previous.accepted ∩ eligible, filtered by
       (in raw_now OR in raw_prev OR NOT adjacent)
new = raw_now ∩ raw_prev ∩ eligible, excluding kept
accepted = first 5 of kept+new, per side
supported = accepted ∩ raw_now ∩ raw_prev
degraded = either accepted side has fewer than 5

9. Complete and partial baskets

Complete accepted 5+5: rerun QP on that membership with current risk even if raw differs. Increases still require supported membership and correctly signed raw weight.

For incomplete membership, unsupported incumbents retain current volume and removed assets get zero. Individually supported assets can receive their shares from the complete current raw QP after full rounding validation. Missing shares are not redistributed (confirmed_raw_targets_v1); actual basket/gross can be smaller.

Raw rounding failure blocks new partial targets but preserves holds and removed-asset exits. A complete accepted-model failure has no partial fallback.

10. Contract sizing and rounding

Last close is reference p. q rounds down to contract step, not coin quantity. Exchange quantity/notional minimums apply; leftovers are not redistributed. ŵ is normalized by G but rounded beta by A. At k=10, 0.03 of A means 0.003 of G.

Full targets require gross 98–100% G, both sides and 15% G cap. Zero insufficient_eligible is exempt. Partial execution need not reach 98% G.

qᵢ=sign(wᵢ) floor_step(|wᵢ|G/(pᵢcᵢ))
ŵᵢ=qᵢpᵢcᵢ/G
0.98≤Σ|ŵᵢ|≤1+10⁻⁸; max|ŵᵢ|≤0.15+10⁻⁹
β_A,target=(G/A)Σŵᵢβᵢ; |β_A,target|≤0.03
exists ŵ>0 AND exists ŵ<0

11. Margin and notional

At E=A=1,000 USDT and k=10: G=10,000, asset cap 1,500; L=20 implies about 500 initial margin plus 20 reserve: 520≤900. This is a calculation, not guaranteed margin availability. Leverage does not multiply G again.

Even empty/partial targets are assessed against the full model: maximum of full-model and saved-contract margin plus reserve. Falling equity does not shrink saved contracts automatically. Missing composition/leverage means unavailable, not zero.

G_used/M_used include positions and remaining opening orders at max(order price,current price). Closing orders release nothing until proven fills. Limits are rechecked at dispatch.

Exchange leverage defaults to 20×, with 10× for AVNT on both sides. Margin is calculated using each asset’s leverage; G/20 is not a valid portfolio estimate when AVNT is included.

R=0.02A; margin_limit=0.90E
M_model=Σ|wᵢG|/Lᵢ; M_contracts=Σ|qᵢpᵢcᵢ|/Lᵢ
max(M_model,M_contracts)+R≤0.90E

Opening order added notional N:
M_used+N/L+R≤0.90E; N/L+R≤available_margin
G_used+N≤min(G_saved,10A_current,10.3E_current)
N_symbol+N≤0.15min(G_saved,10A_current)

12. Differences, deadband and slices

Differences are computed per side. Ordinary nonzero-target adjustment skips a delta below 1% A, checked at current and order price. Zero target, emergency and protective reductions bypass deadband.

Initial slice≤10% A, not G, then limited by gross, cap, margin, beta, step and closeability. To avoid an invalid tiny remainder, the whole position may exceed the usual slice only within required reduction and hard beta. Foreign volume is not adopted; available/frozen are not subtracted twice.

Δq_side=q_target,side−q_current,side
ordinary nonzero target: skip if |Δq|pc<0.01A
q_child≤min(|Δq|,0.10A/(p_order c))
close_capacity=max(0,min(position.available,
                        position.size−remaining_close_orders))

13. Beta and arbitrary partial fills

B_A uses current A, not G. Pending orders may fill independently: buy adds positive, sell negative exposure regardless of open/close. Order valuation uses max(order,current price).

Hard band ±0.10; ordinary slices use ±0.09, reserving 10% for latency/movement. Final guard uses ±0.10. Out-of-band beta may improve without a worse opposite excursion. A pending breach cancels the largest owned contributor to the breached edge, then reconciles.

B_A=Σsign(position)q c mark β/A
δ_o=sign(buy/sell)remaining_o c β max(order_price,current_price)/A
B_low=B_A+Σmin(0,δ_o); B_high=B_A+Σmax(0,δ_o)
L=min(−tolerance,B_A); U=max(tolerance,B_A)
u=|cβ max(order_price,current_price)/A|
room=U−B_high (positive impact), B_low−L (negative)
q_allowed=floor_step(max(0,min(q_requested,room/u)))
existing envelope outside [L,U] → q_allowed=0
u<10⁻¹⁶ → floor_step(q_requested)

14. Passive prices and repricing

b/a are best bid/ask, m midpoint, σ residual hourly volatility. Offset is fixed at transition creation then decays with age. Stale-plan or missing-σ reductions start at zero offset.

Passive entry 14,400s (4h), reduction 7,200s (2h), then IOC may apply. Buy rounds down, sell up to tick. Entry is bounded ±0.3% from current H1 authorization midpoint. Crossing post-only is blocked.

Ordinary reprice only moves toward a fill (buy higher,sell lower), at least 2 ticks and 120s since last submit/reprice. Lost authorization, incompatible target or phase change cancel without waiting.

m=(b+a)/2
o₀=min(0.001,max((a−b)/m,0.10σ))
o(t)=o₀[1−min(1,elapsed/T_passive)]
p_buy=floor_tick[b(1−o(t))]
p_sell=ceil_tick[a(1+o(t))]
entry buy≤m_auth*1.003; entry sell≥m_auth*0.997

15. IOC, authorization and expiry

IOC is a limit order canceling unfilled remainder. Buy bounded by ask×1.001, sell by bid×0.999, with ±0.3% H1 midpoint bound for entry. Rounded buy<ask or sell>bid blocks without market fallback.

At most 3 entry IOC attempts per authorization, at least 120s apart. New confirmed H1 updates midpoint, expiry and attempts, not arrival, offset or age. Entry expires at 86,400s. Mandatory exit/emergency has no total deadline; emergency retries no more often than 10s.

Raw disappearance and renewed confirmation can start a new episode. A material post-expiry target addition permits only the addition; old remainder is carried separately, not rejuvenated. Backward clocks block entry. Unknown submit retains obligations.

IOC buy=floor_tick(ask*1.001)
IOC sell=ceil_tick(bid*0.999)
entry buy=floor_tick(min(ask*1.001,m_auth*1.003))
entry sell=ceil_tick(max(bid*0.999,m_auth*0.997))
entry attempts/authorization≤3; entry age<86400s

16. Freshness, scheduling and targets

New D is created only within [D,D+3600); duplicate processed D is forbidden. Entry expires D+3720; without a new target entries block and resting entries cancel. Permissible exits/protection continue.

Public worker checks overlap and atomically publishes the complete history; exact inputs are SHA-256 archived. Runtime snapshot/book age≤30s and nonnegative. Entry economic statements verified within 120s.

One ordinary order per instrument/side. Queue favors least recently serviced eligible transition, then reduction, smaller resulting |beta|, symbol. Reprices share the queue. Protection compares |beta| first. Ordinary exits do not globally block all openings.

new plan: D≤now<D+3600
entry expiry=D+3600+120
0≤book_age≤30s; 0≤snapshot_age≤30s
normal priority=(last_service,reduce_before_open,|beta_after|,symbol)
protective priority=(|beta_after|,reduce_before_open,symbol)

17. Beta, gross and margin protection

Protection triggers on emergency, |B_A|>0.10, gross excess or M_used+R>90% E. No openings; IOC reductions take priority. Model target remains separate from protective execution.

Gross/margin reductions are proportional; a position worsening excess beta also reduces toward half-band (0.05). Quantity is capped by position then closeability, price and beta envelope. The cause is rechecked at dispatch.

G_limit=min(10.3E,1.03*10A)
z=min(1,G_limit/max(G_used,10⁻¹²),
      max(0,0.90E−R)/max(M_used,10⁻¹²))
q_reduce=q_position*(1−z)
if sign(position)βᵢB_A>0 and |B_A|>0.10:
  q_reduce=max(q_reduce,(|B_A|−0.05)A/(mark*c*|βᵢ|))
emergency: q_reduce=q_position
q_reduce=min(q_position,q_reduce)

18. PnL, funding and cash flows

Realized PnL is reconstructed from confirmed fills with per-side average entry and receipt deduplication. Closing beyond proven inventory blocks accounting. Fees are positive costs, rebates negative. Open PnL uses current positions; signed verified funding and cash flows are separate.

Wallet reconciliation uses equity/unrealized from one balance response, not marks from separate REST reads. Gap>0.01 USDT, unverified fees/stale statements block entry. Deposits are not PnL and do not reset A₀. Turnover is Σqpc over fills. Diagnostic maker/taker follows post-only/IOC types, not guaranteed exchange classification.

avg_new=(q_old*avg_old+q_fill*p_fill)/(q_old+q_fill)
PnL_close=sign(side)q_fill*c*(p_fill−avg_entry)
PnL_open=Σsign(side)q_position*c*(mark−entry_price)
PnL_trade=PnL_realized+PnL_open−fees+funding
expected_equity=E₀+PnL_trade+cashflows
wallet_gap=(E−U_balance)
  −(E₀−U_balance,0+PnL_realized−fees+funding+cashflows)
|wallet_gap|≤0.01 USDT

19. Emergency loss: exact boundary

Trigger: cumulative trading PnL≤−10% A₀. Not trailing or peak drawdown. A₀=1,000 then growth to 1,500 without flows leaves threshold around 900: 40% from peak. Gross 10A means a 1% adverse portfolio move can cost about 10% A before costs.

Emergency persists and sets execution targets to zero; a better score does not cancel it. No total deadline, but price, closeability, beta and retry interval≥10s remain. Local best-effort protection depends on process, network, accounting and liquidity; managed exchange TP/SL is off.

Accounting exceptions make loss checking unavailable that cycle and block entry; independent beta/gross/margin checks continue with a usable snapshot. An unverified report blocks entry but its computed negative PnL can still trigger emergency.

emergency ⇔ PnL_trade≤−0.10A₀
A₀ survives restart/checkpoint/recovery
A₀≠peak equity; A₀≠G
persist emergency; executable_targets←0

20. Journal and blockers

Intent with unique client_order_id is persisted and fsynced before request; target revision, emergency, price, age, capital, beta, margin, ownership and size are then rechecked. Unknown submit/cancel is not blindly repeated and retains reserves. REST orders/positions are not atomic; mismatch does not prove manual closing.

Entry blocks on foreign/unknown orders, ownership mismatch, unreconciled state, wrong account mode, unavailable leading catalog, unconfirmed configured leverage on both sides (AVNT 10×; others 20×), stale data, economics, unsupported candidate, opposite position, margin, lot, price, beta/cap, expiry or backward clock. Entry blockers must not disable permissible owned reductions.

Schema 2 preserves hashes/revisions, membership, age/authorization, economics and A₀. Checkpoint after 128 intents requires no unknown outcomes and a valid receipt checkpoint; obligations persist. Initial baseline needs a clean account/new ledger. Old state requires explicit recovery, no auto-reset.

21. Limits and copy trading costs

No guaranteed returns, liquidation immunity or exact 10% loss ceiling. Historical BTC beta leaves other factors, correlation changes, funding/basis, liquidity and exchange risk. Weights do not optimize turnover; the volume filter is not market capacity. An H1 candle touching price does not prove a limit fill.

Rules/tests do not replace economic validation. Source is not yet public: this reveals formulas and checked algorithms, not an independent source audit. Profile history may include previous versions.

No subscription fee; 25% profit share; fees/funding separate. HTX determines period, sharing base and loss carryforward: 25% of each winning trade would be unjustified without verifying terms. Followers have their own fills/settings.

22. Model and trading parameters

Fractions are decimals: 0.10 = 10%. seconds means seconds; ticks means minimum price increments. max_passive_seconds constrains valid passive-phase configuration rather than adding a separate exit deadline in the loop. membership_confirmation_bars=2 and strict_target_margin_feasibility=true are mandatory validation settings. max_ioc_attempts is retained for the daily path; hourly uses max_ioc_attempts_per_authorization. Live/post-only/IOC flags are operator permissions, not proof of successful execution.

23. Sources and checksums

Hashes identify the checked source snapshot. They are not an audit of exchange execution and expose no credentials or account state.

strategies/rs_rotation_v1/signal_engine.py

1e8ab8ae76f5e6408ec95a83d8f19fd62c6c686d01650fcf0a2ed7bce924a8a8

strategies/rs_rotation_v1/plugin.py

88ee68a3bb86d9785dc233d98f1a91ee7d3c2d85101fcfb018d6931046fd9daa

strategies/rs_rotation_v1/parameters.toml

358014490dd8ac61526898b58933f1df2f34e1499ad98a1295dbb13be7537003

htxbot/portfolio_targets.py

232a1e43ab22309d7e2122de6a7d7759a7d7f93016087b3399f43e908e59654e

htxbot/portfolio_continuous.py

0f0e9bdf9eb802e484cb7a9070219bbdc97dff429d2572c2415be7e8084d0695

htxbot/portfolio_risk.py

29bc694ee279e058a4fb5a8317c94d08a749d997304a6b953b072b01fe2e7b3d

htxbot/portfolio_state.py

7350fec1a3319e679900bcf63d9629e5bd7796abfa7118764026ca224c7e16cb

htxbot/portfolio_engine.py

e54c8653f4efa342e3705cdddc4e8d84bb265df99958c4835fcd8213cabc966f

htxbot/copytrade_execution.py

530903e1684dacb158ac66f72debab06d8f42416ddb30cd352f1de02ddf10b20

htxbot/portfolio_runtime.py

fbb47883114e61679c282b5f25093b404ab9d6991240e1742f39d25adabc4484

htxbot/portfolio_history.py

08730a203490b7e0298eb0c466147eb2b9c9eacb93d602d37719735933d83396

copytrade_config.py

9c9e9ae67e25535f5f63c395b8997ab6bbc0d7debace7412094ec4c585a9bb27

htxbot/secondary_universe.py

838d47a30c91378244d41bbc2f174d5fff1c032a1b6274726c97cd36ab6e686c

config/copytrade.toml

f027564adbe3b7e0680184a0a63f6690bf5a7cd3216cd3b06bb98898360f75fa