fill-real

An execution-grounded dataset for evaluating Solana trading strategies — and evidence that screen-price backtests are systematically biased.

v2.0 — August 2026. 1,746,551 rows, 25× the v1 release. Two new results: an exhaustive condition sweep that returns zero positives, and a paired bound on the horizon-truncation bias that v1 could only flag as unbounded.

Almost every publicly shared backtest of Solana token strategies is computed on screen price: the value an aggregator reports on a polling interval. It is what is available, and it looks reasonable.

The price at which an order actually executes is different. The difference is not symmetric noise that averages out — it is a directional bias, and it is large enough to invert conclusions.

The finding

The same population of tokens, measured both ways:

Measurement	Screen price	Real fill
Migration-speed "ladder" (n≈24k / 21k)	+36%	−69%
Live entry gate cross-check	−42.8%	−1.1%
Live exit policy calibration	−29.8%	−10.9%

The bias runs in both directions, which is what makes it dangerous. Screen price made a ruinous entry rule look profitable. Separately, it made reactive exit policies (stops, trailing stops) look far worse than they are — a snapshot grid fires stops on transient dips that a real order would never have paid.

Screen price penalizes reactive policies and rewards illusory ones.

Pool liquidity does not lead price. Across 150 collapses, the median lead between a liquidity-drop threshold and a −25% price drop is 0 seconds, and at thresholds ≥15%, zero cases had any warning at all. Price is a function of pool reserves — they are the same variable, not two signals. Median single-interval gap: 97.2 percentage points. This closes a whole family of "watch liquidity to exit before the rug" designs.

Real round-trip friction is 10.7% (7.79% entry overhead, n=164; 2.91% exit slippage, n=922), while the median token moves only +8.3% in the best 15-minute window. When friction exceeds the asset's median move, no exit-timing policy can help.

New in v2

1. An exhaustive condition sweep: 0 of 114 have a positive mean

The obvious response to a negative result is "you just haven't found the right filter yet." This tests that claim exhaustively rather than anecdotally.

Over 36,736 real-fill positions, every combination of 10 ex-ante features × 7 percentile cuts × 2 directions, plus 4 social booleans and one cross-strategy condition — 114 distinct conditions with n ≥ 100 each:

best conditions	n	median	mean	win%	ruin%
baseline (all real fills)	36,736	−51.19%	−33.59%	24.4	24.5
preBuyers >= 326	207	−19.82%	−3.25%	34.3	1.9
liqUsd >= 204,183	1,833	+12.53%	−7.33%	67.1	28.8
devBuySol < 0.0395	2,774	−12.78%	−8.14%	45.2	25.7

>>> CONDITIONS WITH POSITIVE MEAN: 0 of 114

A sweep that returns zero positives requires no multiple-comparison correction, because there is nothing to discount. Only ex-ante features are eligible — a condition is a trading rule, so realised hold time, peak multiple and exit reason are excluded by construction, because conditioning on outcomes manufactures edge that cannot be traded.

Two sub-results worth more than the headline:

Win rate and median are actively misleading here. Three conditions win more than half the time and every one of them loses money. liqUsd >= 204,183 wins 67.1% of the time with a +12.53% median — and a −7.33% mean. The left tail is fat enough that a two-thirds win rate is not enough. Report means.

Ruin is predictable; profit is not. preBuyers >= 326 cuts the ruin rate from 24.5% to 1.9% — a genuine, large, reproducible effect. Its mean is still −3.25%. The predictable set and the profitable set are disjoint.

The script that reproduces this ships with the release, inside fill-real-v2.0.0.zip on Zenodo.

2. The horizon-truncation bias, now bounded

v1 shipped a warning it could not quantify: price series truncate near 36 minutes, so "any gate-level figure computed from this data is biased optimistic by an amount that is not yet bounded."

v2 ships 1,728 tokens with a full 190-minute horizon, which bounds it:

horizon	trimmed mean mult	median	share above 1×
36 min (the v1 cutoff)	0.9143	0.2915	34.4%
190 min	0.5970	0.1017	19.3%

The same tokens lose a further 31.7 percentage points between minute 36 and minute 190. 66.3% keep falling after the cutoff.

Read this as a paired estimate, not a universal constant. The comparison is within-token — the same 1,728 mints measured at two horizons — so the drift cannot be a composition artifact. But this subsample was collected 16–21 Aug, and that window is a worse regime than the full population: −19.5 pp at the shared 36-minute mark. The direction is established on paired data; the magnitude is a point estimate from one six-day window. Both figures are published so you can judge it yourself.

Dataset structure

File	Rows	Contents
grad-entry-shape.jsonl.gz	1,602,508	New in v2. Post-migration price trajectories: one row per (mint, seconds since migration, price) across 40,427 mints
liquidity-track.jsonl	65,861	Liquidity, price, FDV and 24h flow series per position
grad-early-shadow.jsonl	38,503	New in v2. Exit-policy grid: one row per (position, exit configuration) with its realised return
grad-social-shadow.jsonl	37,401	Post-migration positions with real Jupiter fills and ex-ante features
exit-slippage.jsonl	922	Real exit slippage, order by order
entry-exec.jsonl	759	Entry overhead: fill vs. decision price
survivor-shadow.jsonl	597	New in v2. Independent survivor-strategy evaluations; joins by mint

1,746,551 rows total, ~74 MB. Collected 13 June – 21 August 2026 from a live system on Solana mainnet. All files are newline-delimited JSON. Row counts and SHA-256 per file are in MANIFEST.json.

Loading

import pandas as pd

# the sweep population
df = pd.read_json(
    "hf://datasets/crdkzk/fill-real/grad-social-shadow.jsonl",
    lines=True,
)

# the price trajectories (gzipped)
shape = pd.read_json(
    "hf://datasets/crdkzk/fill-real/grad-entry-shape.jsonl.gz",
    lines=True,
    compression="gzip",
)

Because the schema varies across the collection period (see limitations), pandas will fill absent columns with NaN. Check for column presence rather than assuming a fixed set.

⚠️ Limitations — read before using

1. Most price series still truncate at ~36 minutes. v2 bounds this bias (see above) but does not remove it: 83.5% of mints truncate near 36 min, only 4.3% reach 190 min. Use the 190-minute subsample to correct, not to replace.

2. The schema changes over time. Older rows lack screenMult and exitImpactPct; newer rows lack devBuySol, migrateDelayMin and the adaptive* fields. Any loader must tolerate missing keys.

3. Single-operator data. Execution telemetry comes from one wallet's order flow. Fill quality may differ at other order sizes. Order sizes and observed price impact are published so transferability can be judged.

4. posSol in exit-slippage.jsonl is mark-to-market value at exit, not entry size. It satisfies slipPct = jupOutSol / posSol − 1. Grouping returns by it measures causation backwards — a position that fell 99% is small because it lost. Join to entry-exec.jsonl (requestedSol, spentReal) for entry size.

Privacy: the dataset contains no wallets, no keys and no transaction signatures. The release is produced by a build script that strips those fields, re-audits its own output, and fails the build rather than shipping a violation. Token mints and creator addresses are public on-chain data.

Not investment advice. Published for research purposes.

Method

The obvious risk in two months of analysis over the same datasets is data fishing: test sixty ideas, keep the one that looks good, never correct for the fifty-nine you discarded. This was treated as a process problem.

• Pre-registration. Every hypothesis is recorded with its prediction and a timestamp before the analysis runs. Only data after that mark counts as validation.
• Negative results are recorded. ~60 hypotheses closed or refuted, with their numbers.
• Artifacts are recorded. One calculation produced means of +19,216% before the reference price was found to be misspecified — written down, with cause and correction.
• Caveats against each verdict are recorded, not just supporting evidence. The regime caveat on the truncation bound above is an example: it weakens the headline, and it is in the headline.

Links

• Code and documentation: https://github.com/cristiandkzk/fill-real
• Archived, citable (concept DOI, always newest): https://doi.org/10.5281/zenodo.21830479

Citation

@dataset{diaz_2026_fillreal,
  author    = {Díaz, Cristian Gonzalo},
  title     = {fill-real: an execution-grounded dataset for
               evaluating Solana trading strategies},
  year      = {2026},
  publisher = {Zenodo},
  version   = {2.0.0},
  doi       = {10.5281/zenodo.22051123},
  url       = {https://doi.org/10.5281/zenodo.22051123}
}

To cite all versions rather than v2 specifically, use the concept DOI 10.5281/zenodo.21830479.

License

CC BY 4.0 — use it freely, credit the source.

Author

Cristian Gonzalo Díaz — @cristiandkzk

Built and operated the instrumented system this data comes from. The methodological discipline is the substance here: this work exists because the measurement kept contradicting the analysis, and the contradictions were recorded instead of discarded.