DIONYSUS allocation strategies in depth

Strategy names

DIONYSUS implements five strategies (enum values in the API). Names match what you send in risk_settings.allocation_strategy:

Strategy	Typical use	Notes
EQUAL_WEIGHT	Equal dollars across ranked symbols (subject to caps)	Ignores softmax, Kelly, and custom stage weights.
CONFIDENCE_WEIGHTED	Tilt toward higher signal / model confidence	Uses softmax over raw confidence weights; softmax_temperature spreads or concentrates the distribution.
STAGE_WEIGHTED	Tilt toward Stage 1–2 vs later-cycle labels	Optional stage_weights per stage string (e.g. STAGE 1 … STAGE 4).
KELLY_CRITERION	Size from Kelly-style fractions derived from modeled edge	half_kelly (default on) and reward_multiplier scale aggressiveness.
HYBRID_WEIGHTED	Combines confidence and stage signals	Uses softmax + stage multipliers; softmax_temperature and stage_weights apply.

These are implementation labels in the product—not a guarantee that any strategy matches academic textbook definitions in every edge case.

Default stage weights (when not customized)

If you do not pass stage_weights, DIONYSUS uses a built-in table aligned with the Weinstein-style stage framework (favor Stage 2, penalize Stage 3–4). Exact numeric defaults can change in the service; treat custom stage_weights as the source of truth for your account when you set them.

Knobs that actually matter per strategy

From the documented API behavior:

• All strategies respect max_risk_per_trade, max_portfolio_risk, max_position_size, max_positions (within the implementation’s greedy sizing and rounding).
• CONFIDENCE_WEIGHTED: softmax_temperature (optional; higher → more uniform weights).
• STAGE_WEIGHTED / HYBRID_WEIGHTED: stage_weights optional.
• KELLY_CRITERION: reward_multiplier, half_kelly; does not use softmax or stage_weights in the same way as hybrid.
• EQUAL_WEIGHT: Ignores softmax, reward_multiplier, stage_weights, half_kelly.

Constraint packs can still merge sector_caps.DEFAULT into max_position_size and rescale softmax temperature when a vol target is set—see Constraint packs and optimizer diagnostics.

Softmax intuition (confidence / hybrid)

Raw scores are converted through a softmax: higher temperature → weights closer to equal; lower temperature → weights more concentrated in top names. This is a mechanical normalization, not a forecast of returns.

Kelly intuition

Kelly-type sizing is sensitive to inputs. DIONYSUS applies caps and minimum position filters so tiny fractions do not create noise positions. Retail users often treat half-Kelly as a starting point; see Allocation strategies for general caveats.

What strategies do not do

• They do not solve a full mean–variance problem with a live covariance matrix (see Σ / covariance placeholders in schema docs—reserved for future work).
• They do not replace tax, liquidity, or ESG judgment—you must still apply those outside the solver.

Related guides

• Portfolio optimization with DIONYSUS
• Allocation strategies — conceptual definitions (Kelly, equal weight, etc.)
• Risk management fundamentals
• How DIONYSUS filters and ranks symbols