Trading strategies

Example and sandbox start: $1,000 nominal USD (100% cash), BTC $0, gold $0, S&P 500 $0. Portfolio size is arbitrary; calculations use weights P_t=[w_USD,w_BTC,w_Gold,w_SPX], sum=1, each in [0,1] before tighter configurable constraints.

Spot/unleveraged/long-only; the historical charts are references, not proof of investable spot instruments. Earlier preference for persistent nonzero class exposure conflicts with the cash-only launch example: treat 100% cash as a launch exception; any subsequent class floor is an unresolved configuration, not an approved 10% rule.

P_(t+1)=P_t+Delta P_t. A sustained signal can move capital repeatedly; one noisy observation should cause little or no movement. Avoid mapping a risk score directly to an exact deployment percentage.

Illustration only: 50/20/10/20 (USD/BTC/gold/SPX) plus a 2-point deployment split 1.2 BTC and 0.8 SPX becomes 48/21.2/10/20.8. Each next action starts from that new state.

Layer 1 decides DeltaRiskExposure: positive transfers USD to deployed BTC+gold+SPX, negative transfers deployed assets to USD. Layer 1 does not choose the recipient.

Layer 2 chooses redistribution among BTC, gold and SPX. A zero-sum rotation can change their mix without changing total deployed capital.

Start 30/30/15/25. Layer 1 deploys 1 point; Layer 2 favors gold over BTC. One illustrative outcome is 29/28/18/25. Dollar-for-dollar transfer, not naive normalization.

Daily ordinary research range: 0–2 percentage points of portfolio value; exceptional 3–4 points is a hypothesis, not policy. Daily candidates: price, momentum, relative strength, volatility, credit, DXY, rates, breadth and shocks.

Weekly ordinary research range: 0–5 points; larger only if tested. Weekly candidates: medium/long momentum, liquidity, macro, credit, valuation, earnings, structural relative strength. Do not double-count the same information in daily and weekly actions. No forced turnover.

Market trend: SPX return; SPX/50DMA and /200DMA; percentage of stocks above 200DMA. Volatility: VIX level/change/term structure and SPX realized volatility. Credit: high-yield and investment-grade spreads, HYG versus Treasuries.

Rates: 10Y and 2Y Treasury yield changes, 10Y real yield, yield curve. Dollar: DXY trend and momentum. Liquidity: Fed balance sheet, bank reserves, Treasury General Account, reverse repo, financial conditions. Macro: PMI/ISM, initial claims, unemployment trend, earnings revisions. Inflation: CPI trend, core PCE trend, breakevens. Cross-asset: BTC trend, copper/gold, stocks/bonds and gold/SPX ratios. Positioning/sentiment/flows: put/call, AAII, fund flows.

All are candidates, not sourced live indicators or hard-coded weights. Yields, inflation and dollar moves need regime interactions: their signs are not uniformly risk-on/off.

Compute 1D, 5D, 20D, 60D, 120D and 252D returns for BTC, gold and SPX, plus volatility-normalized versions. Calendar gaps and BTC weekends require explicit alignment.

Each asset: price/20DMA, /50DMA, /200DMA; 50DMA/200DMA; normalized distance (price−MA)/MA for all three horizons; moving-average slopes and trend acceleration/deceleration. Do not compare raw dollar distances across assets.

Pairs: BTC/SPX, BTC/gold, gold/SPX. Candidate features: pair ratio and momentum, ratio versus MA, 5D/20D/60D/120D relative returns and acceleration.

RS(BTC,SPX)=R_BTC−R_SPX. Compare raw, BTC-vol-normalized and symmetric both-asset-vol normalized formulations; no winner presumed.

Per asset: 10D/20D/60D realized vol, change, downside vol, ATR/price, vol percentile/regime/acceleration. BTC's numerical volatility differs structurally; normalize to prevent spurious dominance.

RAM_i=Return_i/RealizedVol_i over multiple horizons (5D/20D/60D/120D); test Sortino-like and downside-adjusted variants alongside raw momentum, not in place of it.

For BTC/gold/SPX: drawdown from all-time and 20D/60D/120D highs, velocity, days since high and recovery velocity. DD_t=Price_t/RunningMaximum_t−1.

Drawdown alone is not a buy/sell command; it can indicate deterioration, mean reversion, regime transition or noise.

DXY level/return/momentum/50DMA/200DMA/volatility; 10Y real-yield level/change/momentum; 2Y and 10Y nominal yields and changes, 2s10s curve/change; CPI and trend, core CPI, core PCE, inflation surprise, 5Y/10Y breakevens and changes.

Expose shared variables to all three assets; do not force real yields to affect only gold.

Candidates: Fed balance sheet, bank reserves, TGA, reverse repo, financial conditions, money supply and reliable global-liquidity proxies. Test the incremental BTC relationship empirically rather than encoding it.

ISM Manufacturing/Services, PMI and change, initial claims, unemployment, earnings revisions, economic-surprise indices, industrial production and retail-sales trend. Interpret jointly with monetary conditions and publication lags.

High-yield and investment-grade spread levels and daily/weekly changes; HYG trend and HYG/Treasuries relative strength; credit-spread momentum. Credit deterioration can flag a fragile equity rally.

SPX context: percentage above 20DMA/50DMA/200DMA; advance/decline; new highs vs lows; equal-weight versus cap-weight SPX; small-cap/large-cap and cyclical/defensive sector relative strength. Measure concentration versus broad participation.

Spot BTC ETF net flows, spot volume/change, futures basis, funding, open interest/change, liquidations, stablecoin supply/change, realized capitalization, MVRV-like measures, long-term holder behavior, exchange balances/flows.

Derivative observations are information only; strategy remains spot-only. Require incremental predictive value versus price/trend/liquidity before inclusion. Pre-inception ETF flows are unavailable, not zero.

Gold ETF flows, central-bank demand, real yields/momentum, dollar relationship, inflation expectations, gold/silver and gold/copper and gold/SPX ratios, gold volatility, reliable futures positioning.

Central-bank purchases and other slow releases enter only after public availability, not the period described. Historical gold futures chart is not spot execution.

Forward earnings estimates/revisions, forward P/E, earnings yield, equity risk premium, breadth, sector breadth, equal-weight relative strength, credit conditions, margin expectations and earnings-growth expectations. Treat valuation as slower state, not a daily trigger.

Keep four current weights, deployed share, portfolio volatility/drawdown, each concentration, days since meaningful rebalance, recent turnover and daily/weekly actions. Identical market features may warrant different bumps at 95% USD versus 95% deployed.

Portfolio state describes us, not the outside market; separate its schema from market signals.

Expose Delta P_(t−1), Delta P_(t−2) and bounded recent history; 1D/5D/20D turnover; cumulative recent deployment and BTC/gold/SPX rotation. This helps a learning policy avoid ping-pong behavior.

Compare continuous to discrete DeltaRisk actions. An illustrative discrete daily set is −2, −1, −0.5, 0, +0.5, +1, +2 percentage points; zero is first-class. Weekly limits may be wider.

These are research candidates, not optimized policy. A requested action is bounded against actual available USD/deployed balances by the risk controller.

One possible pure-rotation pressure vector A2=[a_BTC,a_Gold,a_SPX] with sum zero. Example [+0.8,−0.5,−0.3] expresses relative direction, not percentages.

Research a constrained mapping into bounded dollar transfers among held assets. Layer 2 can rotate even if Layer 1 chooses zero deployment change.

Given DeltaRisk and the rotation vector, produce Delta P subject to constraints and exact dollar conservation. Funds entering a sleeve come from USD or another sleeve; funds leaving go to a named recipient.

Avoid Normalize(P+ExposureDelta+RotationDelta) if normalization quietly changes exposures the model never requested. Distinguish daily and weekly intents and combine/cap once to prevent double-counting.

After learned or rule-based requests, enforce configurable daily/weekly turnover, per-asset max, optional minimum USD, portfolio volatility, action size, minimum trade size, cost threshold, slippage and cooldown/hysteresis. Invalid actions are corrected or rejected deterministically.

Log requested X and permitted Y separately. The controller prevents pathological actions; it is not another return predictor. No live-trading authority implied.

Expose short/medium/long momentum, trend, relative strength, vol, credit, liquidity and macro separately so models can learn interactions. Agreement may justify larger bounded actions; conflicting signals may favor Delta P≈0.

Do not manufacture one hand-weighted master confidence score or force daily turnover.

Candidate transforms: rolling z-score z=(x−past rolling mean)/(past rolling std), percentile, rolling min/max, vol/rate-of-change, distance from past mean and cross-sectional scaling.

Fit and compute with values available by decision time only; no centered windows, full-sample normalization or future revisions. Keep transform version and fitted history in replay custody.

Retain 1D/5D/20D/60D/120D/252D returns where available instead of assuming a single horizon. Control feature explosion and publish importance and family/horizon ablation results.

Every feature must record observation period, publication timestamp, earliest usable/tradable timestamp, revision policy, source and frequency. A value enters a decision only after publication and realistic ingestion delay.

August CPI released in September is unavailable in August. Same rule for employment, GDP, earnings, ETF flows, central-bank gold and positioning; use vintages where possible.

Research instruments may be BTC spot, S&P index, gold reference, USD/rates; execution requires separately verified spot BTC, gold-eligible exposure, S&P-eligible exposure and cash/T-bill proxy. SPY/GLD/IAU are ETF candidates, not the index or physical gold; gold futures are not spot.

Respect inception, rights, fees, liquidity, issuer/custody and Hyperliquid eligibility. No claim that all four are currently purchasable as intended on Hyperliquid spot. Never backfill a later-listed asset.

Net return=gross return−fees−spread−slippage, with instrument-specific assumptions and actual filled size. Measure daily/monthly/annual turnover, BTC costs, ETF spreads and fees, trade counts and average sizes.

Small incremental trades can lose their entire apparent edge to costs. Report pre- and post-cost comparisons; do not apply today's fees to history without labeling assumptions.

Compare raw portfolio return against risk-aware reward families: return−lambda_cost·cost−lambda_drawdown·drawdown penalty−lambda_turnover·turnover−lambda_risk·risk penalty. Also explore realized/downside volatility, concentration and tail loss.

Lambdas and training horizon are research parameters. Compare RL to simpler baselines before claiming added value.

Compare 100% USD, 100% SPX, 100% BTC, 100% gold, equal-weight BTC/gold/SPX, equal-weight all four, periodically rebalanced equal-weight, simple momentum rotation, trend-following rotation, volatility-scaled momentum and rule-based incremental rotation.

Use the same dated reference/execution assumptions, costs and out-of-sample windows. The chart snapshot is NOT a backtest of this strategy.

Split training, validation and untouched out-of-sample testing chronologically. Roll forward, retraining only on data and vintages available at each historical decision; freeze policy and hyperparameters before each test segment.

Do not optimize on the entire 15-year history and present its resulting return as out-of-sample. Account for overlapping labels and multiple testing.

Post-hoc slices: equity/BTC bull and bear, inflation shocks, disinflation/deflation, tightening/easing, recession, credit stress, VIX shocks and strong/weak dollar. Compare both portfolio performance and action behavior.

Future-defined regime labels are evaluation lenses, not permissible live model inputs unless defined causally at decision time.

Report CAGR, total return, annualized vol, Sharpe, Sortino, max drawdown, Calmar, worst day/week/month and recovery duration. Report turnover, costs, average weights for all four assets, max concentration and time primarily risk-on/defensive.

Plot allocation and trades through time with USD/BTC/gold/SPX weights, not just ending equity. Define cash yield, dividend, financing, fee and price-vs-total-return assumptions.

For every decision retain raw market and macro inputs, normalized causal features, Layer 1/2 observations and requested actions, pre-action portfolio, deterministic-controller modifications, trades, costs, post-action state and later reward.

Replay should answer why, when and how a 1.2-point deployment happened. Log model/feature versions, source vintages, timestamps and any unavailable inputs.

Ingestion → point-in-time store → market/macro/liquidity/cross-asset/asset features → causal normalization → observation(market signals, portfolio state, action history) → Layer 1 risk + Layer 2 rotation → action combiner → Layer 3 risk/execution controller → simulated trades → next portfolio → reward.

Swappable components. This is proposed architecture, not running services.

Market signals describe outside conditions (momentum, VIX, DXY, yields, liquidity, credit). Portfolio state describes our weights, volatility and drawdown. Action history describes our recent bumps and turnover.

Maintain separate schema groups and timing rules; state/history must not be casually labeled as exogenous market features.

Day 0: USD100/BTC0/gold0/SPX0. Day 1: +1.0 risk point routed to SPX → USD99/SPX1. Day 2: +1.5 with BTC favored → USD97.5/BTC1/SPX1.5. Day 3: no-op holds that state. Day 4 and later: persistent positives can gradually approach illustrative USD30/BTC25/gold10/SPX35 over many observations, not four days.

Deterioration can raise USD 30→33→35→39→43 rather than jump to 100, except for explicitly designed and tested emergency controls. Examples are illustrative, not trained decisions or promised performance.

Given causal conditions, current portfolio and recent actions, what small adjustment improves expected future risk-adjusted outcomes? Not: what is today's perfect complete portfolio? Inertia and bounded transfers define the project.

Phase 1 point-in-time BTC/SPX/gold/USD/rates/vol/credit/dollar/liquidity/macro data and timing contracts. Phase 2 returns/trend/RS/vol/drawdown/ratio and macro transformations. Phase 3 thoroughly tested stateful simulator with requested bump → constraints → fills/costs → next state.

Phase 4 strong rule-based baselines. Phase 5 Layer 1 learning. Phase 6 Layer 2 learning. Phase 7 combined model through deterministic controller. Phase 8 walk-forward out-of-sample evaluation. No RL before causally correct simulator/baselines.

1 no lookahead; 2 exact accounting; 3 reproducibility; 4 point-in-time correctness; 5 realistic costs; 6 strong baselines; 7 modularity; 8 explainability; 9 RL sophistication. A correct boring simulator beats a sophisticated contaminated model.

Deliverable register: (1) technical architecture; (2) indicator-by-indicator data-source plan; (3) canonical feature schema; (4) frequency/timestamp specification; (5) portfolio-state schema; (6) Layer 1 observation/action; (7) Layer 2 observation/action; (8) daily/weekly combination; (9) deterministic controller; (10) simulator; (11) transaction-cost model; (12) baselines; (13) RL environment; (14) rewards; (15) train/validation/test; (16) walk-forward; (17) backtest reporting; (18) ablation; (19) explainability log; (20) lookahead/leakage tests.

Items 1–9 and 11–20 are specified or sketched here as research contracts, not implemented or validated; item 10, the simulator, is NOT built by this desk. First executable milestone, after separate bounded approval: causally correct, stateful simulator replaying daily/weekly incremental bumps exactly with costs and constraints, then baseline comparison, then learning.