What this demonstration establishes
You can change a synthetic shock, portfolio allocation, market depth and participant switches, then inspect a recomputed portfolio path. The same configuration and seed produce the same result.
The browser engine was built specifically for this experience. It is separate from the earlier MicroWorld/Reflex research prototypes. It demonstrates an inspectable workflow and a proposed feedback mechanism, not empirical forecast accuracy.
From state to action to next state
The model starts with a market index of 100, annualized volatility of 12%, full exposure for each participant group and a portfolio value of 100. Each subsequent step reads only the previous simulated state.
1. Participants choose target exposures
- Vol-control: target exposure is min(1, 12% / lagged annualized volatility). The group moves 55% of the way to that target each step.
- Trend followers: target exposure is clamp(1 + 12 × the sum of the previous three step returns, 0, 1). The group moves 40% of the way to target.
- Market makers: target exposure is clamp(1 − 6 × max(0, lagged annualized volatility − 15%), 0.3, 1), an inventory-risk limit. The group moves 50% of the way to target each step.
- Passive funds & ETFs: target exposure is clamp(1 + 4 × previous-step return − 3 × max(0, lagged drawdown − 2%), 0, 1), representing leveraged-ETF rebalancing and redemptions. The group moves 80% of the way to target.
- Leveraged funds: target exposure falls from 1 to 0.5 when lagged market drawdown exceeds 3.5%. It returns to 1 below that threshold. The group moves 65% of the way to target.
All group exposures remain between zero and one. These simplified rules do not represent a specific fund, a complete CTA model or a full leverage/margin system. The trend warm-up returns are +0.4%, +0.3% and +0.2%.
2. Actions meet available depth
effective depth = selected depth / [1 + 2 × max(0, lagged volatility − 0.12)]
weighted flow = 0.30 × Δvol + 0.20 × Δtrend + 0.15 × Δmarket makers + 0.15 × Δpassive + 0.20 × Δleveraged
price impact = 0.12 × constrained participation × weighted flow / effective depth
Depth and participation are expressed as ratios. Negative exposure changes produce selling pressure. Disabled groups contribute zero flow; their weights are not redistributed. Recovery can produce buying pressure later in the path.
3. The market moves
The market return equals the external return plus price impact. External returns follow the selected ten-step shock profile, with a small shared seeded perturbation. Volatility updates as variance = 0.82 × previous variance + 0.18 × return², annualized using 252 steps.
4. Portfolio exposures transmit the move
Growth equities receive 1.35 × market return. Broad equities receive the market return. Duration bonds receive scenario-specific exposure to the external return plus 0.08 × participant price impact. Cash returns zero.
Each asset value compounds separately. The portfolio is the sum of those values weighted by the initial allocation. It is buy-and-hold; weights drift and there is no trading recommendation or automatic rebalancing. Output percentage-point differences use the same portfolio and starting value.
Explicit assumptions
| Input | Default / range | Interpretation |
|---|---|---|
| Initial equity shock | 4%; 0–10% | Sets the negative initial move and scales the external profile |
| Depth | 75%; 25–150% | Relative depth; 100% is the model’s reference, not an observed market reading |
| Constrained participation | 60%; 0–100% | Scales the price impact of the five modeled groups; not a measured market share |
| Initial portfolio | 65 / 25 / 10 / 0% | Growth / broad equities / duration bonds / cash |
| Seed | 42; integers 1–999999 | Fixes the shared pseudo-random external perturbation |
| Event display threshold | 0.6 percentage points | Only exposure changes larger than this are shown as events; smaller flows still affect calculations |
External profiles
At each step, external return = −shock × profile value + seeded perturbation. Perturbations are uniform in [−0.0005, +0.0005) × shock / 4%, so zero shock produces no perturbation. None of these profiles is fitted to a real event.
| Scenario | Ten-step profile | Bond external beta |
|---|---|---|
| Equity repricing | 1, .28, −.12, .08, 0, −.06, 0, −.04, 0, 0 | −0.12 |
| Liquidity squeeze | 1, .35, .15, −.12, .04, 0, −.05, 0, 0, −.03 | +0.18 |
| Rates repricing | 1, .30, .15, −.15, .05, −.05, 0, 0, −.03, 0 | +0.60 |
The model uses Mulberry32 seeded pseudo-random draws. The engine, parameters and complete run can be inspected directly: engine source or export a run from the Lab.
What the comparisons mean
Shock-only path: the same initial holdings, external returns and asset mapping, with all participant price impact removed. It is a model ablation, not a competitive forecasting benchmark.
Sensitivity range: rerun the active rules with the 0.12 impact coefficient multiplied by 0.75 and 1.25. The displayed range spans those two portfolio paths at each step. It is not a calibrated confidence interval, a quantile estimate or an exhaustive uncertainty bound.
Feedback difference: final portfolio return with feedback minus final return without feedback, in percentage points. It is not alpha, avoided loss or evidence that a trader could have achieved a benefit.
Export: a local JSON file contains the engine version, input assumptions, all path values, events, sensitivity runs and limitations. The simulation requires no account, holdings upload or external API call. The site uses no analytics or tracking code.
Where the model stops
- There is no fitted observation model, inferred real fund inventory, order book, margin financing or transaction-cost model.
- This browser demonstration exposes one interactive switch for each of the five participant types, with illustrative, uncalibrated weights. The simplified browser rules omit many real mechanisms.
- Prices alone cannot uniquely identify every behavioral parameter. Independent flow or positioning evidence would be needed for stronger claims.
- Asset exposures and the volatility/liquidity relationship are fixed assumptions. They can dominate the result.
- The engine rejects any active, baseline or sensitivity path with an absolute one-step market move above 35%. This is an explicit demonstration-domain guard, not an empirically validated safety threshold. The interface withholds all results for that configuration.
- The ten-step horizon and bounded controls are for exploration; they do not establish real-time reliability or tail-risk calibration.
- Turning a participant off is an intervention inside this model. It does not identify that participant’s causal contribution in a real event.
- All research extensions, paid workflows and production API capabilities described elsewhere are planned, not delivered by this demo.
The next study needs licensed time-aligned data, predeclared targets, credible baselines, held-out time and sectors, ablations, uncertainty analysis and external review. See the research direction →
Evidence and further reading
These sources frame the problem and existing alternatives. They do not endorse MicroWorld or validate this engine.
- Bank of England, 2024 SWES final report — institutional interactions and stress amplification.
- MSCI Crowding Solutions — established crowding analytics.
- Simudyne Pulse methodology — an existing market simulation approach.
- ABIDES research framework — open-source agent-based market simulation.
Source review: 1 October 2026. Model version: mw-lab-1.1.0. No external source code or customer data is embedded in this demonstration.