Studio Strategies, Part 3: Build More Advanced Experiments

Add complexity only when it answers a question

You can now build a graph and read a replay. The next step is not to attach every available block. It is to identify a weakness in an idea and add a rule with a clear job. Does an entry need confirmation? Does a resting quote become stale? Does another category require different data and an entirely different exit?

The recipes below are research starting points. Their thresholds are illustrative, and feed access depends on your account and selected universe. Each example keeps the graph small enough to explain. Test them separately before combining them.

Experiment A: offer a price instead of taking one

Part 1 used IOC. A maker variation asks whether waiting on the book produces a better entry after accounting for missed fills and adverse selection—the risk of getting filled just as the value moves against you. Start with BTC 15m and preserve the basic idea of a model-estimated price gap.

Entry recipe: Market time remaining > 120 seconds → Random-walk mispricing ≥ 0.05, using 5m volatility history and Best bid → bid price filter from 0.10 to 0.90 → Post-only at zero tick offset, with a 120-second lifetime. Every false branch ends in Hold. The favorable gap is now measured against a passive bid, so this is a different execution experiment from an ask-based IOC.

For example, with bid 0.58, ask 0.60, and estimated value 0.66, the raw gap to the bid is 0.08. A bid at 0.58 may never fill. If the model later falls to 0.55, that previously attractive bid is now above the estimate. Maker fees or possible rebates do not fix that problem.

Give every resting order a cancellation plan

Build Cancel as a chain of checks: time remaining ≤ 120 seconds; otherwise favorable edge < 0.03; otherwise Order age ≥ 30 seconds. Each true branch leads to Cancel, and the final false branch leads to Hold. Use the same bid comparison and 5m history as Entry. Three checks plus an action fit within the five-node path limit.

The 0.05 entry and 0.03 cancellation levels leave a small buffer, so tiny movements do not immediately reverse the decision. The 30-second cancellation rule normally acts before the 120-second order lifetime; the lifetime remains a separate expiry bound. Cancel requests still take time, and a fill can happen before cancellation completes.

For this experiment, keep Exit on Hold and study settlement outcomes, just as in Part 1. You have changed order execution and management, not invented a stop-loss. Compare fills, net P&L, and adverse outcomes with the IOC baseline. Order price distance is another possible cancellation input, but introduce it as a later, separate experiment.

Experiment B: require activity to confirm momentum

Momentum asks whether price has recently moved in a direction. Relative traded volume asks whether the move happened during unusually active trading. Neither tells you that an outcome is cheaply priced, so retain a value check rather than buying activity alone.

Use BTC 15m. Entry recipe: time remaining > 120 seconds → Price momentum on Binance spot, 5,000ms lookback, ≥ 5 basis points toward the evaluated outcome → Relative traded volume on Binance trades, 5,000ms signal window and 600,000ms baseline, ≥ 2 → Random-walk mispricing against Best ask with 5m volatility history, ≥ 0.05 → IOC, zero tick offset. All false branches Hold; Cancel and Exit Hold. Four conditions plus the action use the full five-node path allowance.

Suppose the earlier ten-minute baseline implies $100,000 of trades for an ordinary five-second window. If the current five seconds contain $250,000, relative volume is 2.5 and passes the example threshold. On prediction markets, momentum is oriented toward the outcome being evaluated; on perpetuals, its sign remains the price direction.

Choose a confirmation block for its meaning

Relative traded volume measures executed activity, not the amount of liquidity waiting on the book. Book imbalance measures displayed buying versus selling depth. Taker flow imbalance measures aggressive executed buying versus selling. These are different observations, even when they all appear bullish.

If buyer-initiated notional is $70,000 and seller-initiated notional is $30,000, the raw taker-flow imbalance is (70,000 − 30,000) / 100,000 = 0.40. That is a description of recent trades, not a 40% probability of profit. Test it as a replacement for a confirmation rule, rather than stacking correlated filters until almost nothing trades.

For Experiment B, compare the value-only baseline, then momentum plus value, then momentum plus volume plus value. Keep execution and size constant. If the extra block contributes nothing on fresh periods, remove it. Simpler logic is easier to diagnose.

Experiment C: compare a weather range with a forecast

Weather markets require a different question: what exactly does this contract settle on? Check the city, local date, observation source, daily high or low, units, and rounding. A forecast for a nearby airport or the wrong day is not automatically the settlement measurement.

The professional NBM range edge block supports a narrower case: US daily maximum-temperature YES ranges in Fahrenheit, with both range boundaries inside the supported NBM percentile anchors. It compares an interpolated forecast probability with the normalized midpoint probability across sibling YES outcomes. It rejects unsupported tails, Celsius ranges, and missing sibling books.

Example: a supported bounded range has an estimated probability of 0.32 and a normalized market probability of 0.25, producing a 0.07 model gap. That passes a 0.05 edge threshold. If the actual ask is 0.34, however, this does not establish an attractive immediate purchase. The block’s comparison is not the executable ask and does not subtract fees.

Build the weather experiment around those limits

Choose Prediction markets → Weather, a supported US city, daily maximum temperature, and a nearby event date offered in the universe controls. Entry: NBM range edge with minimum edge 0.05 → current bid price filter from 0.02 to 0.90 → Post-only at zero tick offset with a 120-second lifetime. Route false branches to Hold. Cancel when the NBM edge condition no longer passes, or when Order age reaches 120 seconds. Exit holds to settlement.

This graph tests a passive forecast-value hypothesis; it does not automatically prove the posted bid is below forecast value. Inspect the modeled probability, sibling normalization, and executable quote when interpreting it. Compare one city and date family before broadening the universe.

Weather provider consensus can describe agreement or dispersion among providers, while weather metrics and forecast-change blocks answer other forecast questions. Agreement is not certainty: providers can share inputs and make the same error. Weather evaluation follows a minute cadence, and replay needs the forecasts that were available at the time, not today’s revised forecast for yesterday.

Experiment D: a perpetual-market mean-reversion sketch

Perpetuals have no binary $1 settlement. You must define long and short entries, position-closing exits, and the venue. Mean reversion is the hypothesis that a recent move will partly reverse; it is different from following momentum. Persistent trends can make it lose repeatedly.

A concrete starting sketch for BTC on Hyperliquid: require a spread of at most 8 basis points. If signed ten-minute Price momentum is ≤ −20 basis points, consider a passive buy; if it is ≥ +20 basis points, consider a passive sell. Use a one-tick offset and cancel unfilled quotes after 30 seconds. Otherwise Hold. Treat these as two branches, not two conditions that must both be true.

Define both exits before running: a long closes through a sell marked reduce-only, and a short through a buy marked reduce-only. An illustrative reversal rule closes the long after ten-minute momentum reaches +20 basis points and closes the short after it reaches −20. This rule has no guaranteed maximum loss and can leave a position open through a trend. Review capital, leverage, funding, and separate account loss controls. Do not copy the hold-to-settlement exit from the prediction-market example.

Execution algorithms answer “how much, when, and how?”

TWAP spreads a bounded parent quantity across time. For example, a 100-unit parent with ten equal scheduled slices targets 10 units per slice before accounting for unfilled amounts and venue minimums. VWAP follows a volume profile; the supported implementation needs historical trades to build its pre-start profile. Neither algorithm decides whether the asset is good value.

Pegged orders follow a reference price with cancel-and-replace behavior. IOC retry tries the remaining quantity within its configured limits. More attempts do not guarantee liquidity. These tools are venue-specific; choose from the actions supported by the selected universe, and keep size limits and cancellation behavior explicit. Learn a single IOC or Post-only action before managing a sliced parent order.

Your next three experiments

Pick one recipe, state its reason in a sentence, and save a baseline. Change one rule, record the exact replay window, and compare costs, fills, and drawdown. Then freeze the chosen version and evaluate later non-overlapping periods. Keep unsuccessful versions in the notebook so you can see how much searching produced the apparent winner.

Think about shared exposure before running several strategies. BTC momentum and a BTC prediction-market value buyer can both suffer from the same market move. Different graphs are not necessarily diversified positions. A smaller, understandable experiment is a better learning unit than a large collection of unexplained signals.

Finish the series with three things you can show another person: a graph you can narrate from Entry through Cancel and Exit; a replay record with its assumptions and actual time window; and a written explanation of when the idea fails. That is a useful foundation for further research, whether the first backtest makes money or not.