Robinhood stock-token grid research — September 14, 2026

Open the interactive stock research page for token selection, New York session overlays, grid levels, fills, equity and configuration comparisons. Download the market ranking CSV; it contains all 334 researched contracts with unknown values retained, including the ranked top 100 below. The saved research summary, grid results CSV, and prices with New York timestamps preserve exact numbers.

Research generated 2026-09-14T04:54:07.018726+00:00; volume snapshot spans 2026-09-14T04:13:09.758028+00:00–2026-09-14T04:14:33.329649+00:00 (approximately 00:13–00:14 EDT on September 14). These are historical research snapshots, not live recommendations.

What was tested

The universe begins with 194 official Robinhood contracts plus 59 additional CoinGecko Robinhood listings and 81 Anchored products. We requested every contract from DEX Screener and matched exact addresses on Robinhood Chain. Twelve batch requests and 334 per-token requests returned 2,268 distinct pools; 192 contracts had positive reported 24-hour volume. The top 100 are all official Robinhood tokens. BND and SATS had no returned pools; neither did the other 140 additional listings. Absence from this provider is not proof of no onchain trading. DEX Screener API reference, Robinhood asset API documentation.

Ranking sums reported pool volume after deduplicating chain/pool identities. A stock–stock pool can count toward both constituent rankings, so token totals cannot be summed as unique market volume. Liquidity is independently reported pool value, not executable depth at a chosen slippage. Forty-three individual responses returned at least 30 pools and may be capped; pools found through another token can make aggregate pair counts exceed 30. The ranking covers returned pools, not every venue.

Each asset uses one hourly historical source, with priority DeFiLlama CoinGecko ID, DeFiLlama Robinhood contract, then direct CoinGecko. Providers are not stitched together, and underlying-stock history is not substituted. The research uses sampled reference prices, potentially aggregated across venues/chains; they are not historical bid/ask execution quotes. Current-volume selection introduces historical selection and survivorship bias.

The 144 candidates combine four spacings (0.5%, 1%, 2%, 4%), three level counts (2, 4, 8 per side), three recenter policies (fixed, 24 hours, 72 hours), and four session modes. Each token has chronological elapsed-time splits of 50% training, 25% validation and 25% heldout test. Both the primary configuration and each mode-specific alternative are chosen using training/validation only. Portfolios restart with $10,000 at each split; displayed split returns are not a compounded continuous account.

When a mode is eligible, initial capital is 50% cash and 50% token with purchase costs paid. A timed mode that starts outside its window holds 100% cash until its first eligible observation; it does not buy and immediately flatten. Trading is long-only, cash-constrained and inventory-constrained. Base execution costs are 0.20% per side (10 bp fee plus 10 bp adverse execution allowance); stress costs are 0.50% per side. Entry, session exit and terminal liquidation costs are included. These are assumed costs, not measured venue spreads or fees. The $10,000 model normally begins with a $5,000 purchase, but rank-100 UMC reports only $1,503.85 in total trailing-24-hour volume. Membership in the top 100 therefore does not establish sufficient liquidity for this model notional. There is no historical depth, participation-rate, capacity or venue-fee validation.

A previously standing crossed limit may fill once per observed interval of at most two hours. Replacement orders become active after the observation; recentering follows processing of standing orders. Longer gaps rebuild the grid without inferred fills while inventory stays exposed. This restricted model does not infer intrahour highs/lows and is not a guaranteed pessimistic return bound: omitted purchases can avoid losses.

UTC, New York time, and US sessions

Timestamps remain absolute UTC observations and are converted with America/New_York, including daylight saving changes. During this summer sample, New York is EDT (UTC−4): 13:30 UTC is 09:30 EDT, and 20:00 UTC is 16:00 EDT. In winter EST (UTC−5), those core boundaries become 14:30 and 21:00 UTC. The calendar models NYSE core sessions, weekends, published holidays and early closes for 2026–2028. NYSE calendar.

Premarket 04:00–09:30 and postmarket core-close–20:00 are descriptive display bands, not assertions that a specific token venue supports those hours. Tokens can trade while US equities are closed. Session entries and exits occur at the first actual sample; a closed-session strategy can therefore realize opening-session moves before it is flattened. Missing samples can extend that delay. “Weeknight closures” must not be read as precisely isolated overnight returns.

Price-quality screen

A posthoc diagnostic flags 66 of 100 series: it detects 301 adjacent price moves of at least 25% within two hours and 25 reverting spikes with a center deviation of at least 50%, neighboring prices within 10%, and a neighbor span of at most six hours. These are unexplained observations, not proven data errors. Raw samples and trades remain unchanged; the diagnostic is a reporting gate, never a hindsight fill filter. Passing it does not establish that a quote was executable. Thirty-four series are unflagged; their primary results have 12 positive base-cost returns and eight positive stress-cost returns. No series passes the complete robustness screen.

Heldout outcomes

Across the 100 primary configurations, median heldout return was -1.62% at base costs and -3.07% under stress. Training/validation selected 58 all-hours, 24 core-session, 12 weeknight-closure and six all-closure configurations. Positive absolute return alone does not show a grid advantage over holding.

The following aggregates use the independently selected best training/validation configuration within each mode for each of the 100 assets. They compare 400 exploratory alternatives; they are not four live portfolio returns. Different tokens have different history and heldout dates. Counts and medians retain anomalous source observations rather than silently removing them.

There is no general profitable overnight result here: only nine weeknight alternatives and six all-closure alternatives remain positive under stress; both median stress returns are negative.

Positive primary examples, with their failures

These six examples are ordered retrospectively by heldout stress return. Their configurations were selected before test evaluation; choosing these examples after viewing test results is retrospective reporting, not a new unbiased strategy-selection exercise. Returns are net percentages of initial split capital, not annualized. The matched-hold column is the session-matched 50/50 benchmark at base costs; drawdown is observed test-equity drawdown, also at base costs. Configurations show spacing / levels per side / recenter period.

• USO failed: hourly coverage; validation stress net/excess; continuous 50/50 and full-hold comparison; development weekly consistency; neighboring parameter stability.
• BE failed: unexplained price discontinuities; session-matched hold comparison; continuous 50/50 and full-hold comparison; heldout weekly consistency.
• DELL failed: unexplained price discontinuities; session-matched hold comparison; continuous 50/50 and full-hold comparison; development weekly consistency; heldout weekly consistency; neighboring parameter stability.
• AMD failed: hourly coverage; minimum heldout grid fills; session-matched hold comparison; continuous 50/50 and full-hold comparison; neighboring parameter stability.
• ORCL failed: hourly coverage; neighboring parameter stability.
• RIVN failed: unexplained price discontinuities; hourly coverage; development weekly consistency.

The closest illustrative cases still fail: ORCL lacks sufficient coverage and neighboring-parameter stability; RIVN has unexplained price discontinuities, insufficient coverage and weak development-week consistency. BE and DELL are profitable but their matched holding benchmarks did better. No example passes all screens.

Positive closed-session alternatives

All six examples below belong to price-quality-flagged series. Their positive returns are unreliable modeled outputs, not credible evidence of profitable overnight trading. These are mode-specific alternatives, not necessarily the primary selected configuration. They are also ranked retrospectively within each mode by test stress return and are not validated recommendations.

Reference-price anomalies matter. The core-session alternatives include extreme outputs: UMC +412.05% base/+394.85% stress, CCL +339.08%/+325.73%, and SMCI +239.52%/+229.54%. Their traces include session exits at approximately $188.65 after UMC traded around $20.18, $219.52 after CCL entered at $24.17, and $216.42 after SMCI entered at $37.12. The session-matched holding benchmarks also exhibit enormous gains. These are anomalous reference-price-dependent results, not evidence of executable arbitrage. They remain visible for audit; the counts above must be interpreted with this data-quality limitation.

The heuristic screen requires no flagged unexplained price discontinuities, at least 28 days of history, seven days each in validation/test, 70% hourly coverage in every split, minimum grid-fill counts, positive base/stress test returns, benchmark outperformance, development and heldout weekly consistency, and neighboring-parameter stability. Weekly folds restart funded and are fixed-configuration evaluations, not independent selection trials. Searching 144 candidates across 100 current constituents has no multiple-testing-adjusted significance claim. Passing these gates would still not validate liquidity, spreads, fees, fills or executable profitability; zero configurations pass them here.

Top 100 by reported Robinhood-chain trailing-24-hour volume

USD values are the saved September 14 snapshot, rounded to cents. “Cap?” marks a per-token response with at least 30 pools; “—” means the threshold was not reached, not proof of exhaustive coverage. A liquidity value with † omits one or more returned pools whose liquidity was not reported. Ranking is by volume, not liquidity.

To reproduce the saved research with the existing market snapshot and price files, run python scripts/search-stock-grids.py --workers 8 from the repository root. Refreshing prices, constituents, calendar or code creates a different experiment; preserve the original snapshot when comparing results.