Position-correct Solana wallet win rate: a 194,610-wallet study

In an August 2026 temporal study of 194,610 eligible Solana wallets, wallets in the top decile of formation-period closed-FIFO-lot win rate had positive cumulative gross realized PnL on new lots entered and closed in the following 13-day period 60.70% of the time. The bottom decile did so 13.44% of the time—a 47.25-percentage-point separation.

That is a strong association, but not proof that position-correct win rate is a uniquely better predictor. The side-by-side exit-row sensitivity check was also strong: 64.73% for its top decile versus 14.08% for its bottom decile. In this window, both historical rates separated the closed-only future outcome. The advantage of position-correct accounting is more fundamental: it answers a wallet-level question without turning several profitable exit rows from one entry lot into several independently won positions.

The failure mode can be extreme. During a May 30, 2026 internal endpoint check, one wallet showed a 44.57% positive-exit-row rate and a 1.96% positive-closed-FIFO-lot rate. The former was 22.7 times the latter. That is one dated wallet observation, not an estimate of average market-wide inflation and not a claim about every competing API.

This report publishes the current study design, cohort counts, confidence intervals, side-by-side sensitivity result, limitations, and a machine-readable aggregate result. It deliberately separates those results from older internal notes whose original artifacts were not retained.

Answer first#

• A wallet exit row is not an entire economic position. A single sell can match FIFO lots, and one lot can be realized through several exit rows.
• Conyr keeps two accounting lanes. The exit-row lane counts positive and negative realized slices; the position-correct lane gives a FIFO entry lot one cumulative outcome when it fully closes.
• The current temporal result is large but conditional. It includes wallets with at least 10 decided closed lots in formation and at least 3 decided FIFO entry lots both entered and fully closed in evaluation, so it does not describe all observed wallets.
• Both rates were associated with future closed-lot gross realized profitability. Evaluation lots had to be both entered and fully closed after the temporal split. The study does not claim the closed-lot rate out-predicted the exit-row rate in this window.
• The 22.7× observation is one historical case. It demonstrates severity, not prevalence.
• Open and abandoned bags remain a major limitation. The outcome is cumulative gross realized PnL on closed FIFO lots, not marked-to-market or fee-net wallet return.

The unit-of-analysis problem#

Suppose one FIFO entry lot is sold in four profitable pieces while another entry lot closes at a loss in one row.

Nothing about the fills changed. Only the unit being counted changed.

Exit-row accounting is useful for execution questions: how often was an individual realized slice profitable, how consistently did a wallet take gains, or how much volume left? It is not a clean answer to “how often did this wallet's completed entry lots work?”

The distortion does not require bad source data. It appears naturally when winning lots are sold in more pieces than losing lots, or when exit style varies across wallets. A leaderboard can faithfully total every row and still let exit frequency masquerade as position-level skill.

What “position-correct” means in Conyr#

Conyr's production accounting is FIFO-lot based. Each qualifying acquisition creates an entry lot. A disposal can match one or more lots, and the accounting pipeline emits a realized-PnL row for each matched slice.

Let a decided exit row be a realized slice with PnL above or below zero. Breakeven rows are retained but excluded from the rate denominator:

exit-row win rate = positive realized exit rows
                    ---------------------------
                    positive + negative exit rows

Let a decided closed lot be a FIFO entry lot whose cumulative gross realized PnL is above or below zero when the lot reaches its full-close state:

closed-lot win rate = profitable fully closed FIFO lots
                      ---------------------------------
                      profitable + losing closed FIFO lots

The lot identity includes token mint, entry transaction signature, and entry event index. The event index matters because one transaction can contain multiple acquisitions that must not collapse into one lot. Partial realizations update the lot, but do not create extra closed-lot wins. Cumulative lot PnL supplies the one final outcome when its remaining balance reaches zero.

In the current production field, that cumulative lot value is gross of the separately allocated entry and exit network-fee fields. The study follows the production outcome semantics so its rates map to the API, but its dollar outcome is not net profitability.

The public API exposes these lanes as:

• raw_exit_win_rate_pct: “What share of decided realized exit rows made money?”
• position_win_rate_pct: “What share of decided fully closed FIFO entry lots made money?”
• win_rate_inflation_pts: the non-negative difference raw exit rate − closed-lot rate.

The API keeps the familiar word “position” in its field names. The exact accounting grain is a fully closed FIFO entry lot, which is not necessarily the wallet's entire economic exposure to a token when it has multiple acquisitions.

Neither rate is a probability that the next trade will win. Both summarize observed history.

The August 2026 temporal study#

The current study freezes two adjacent 13-day UTC windows from the post-wipe streaming production corpus.

After applying those rules, 194,610 wallets were eligible. Across the full eligible cohort, 58,411 wallets, or 30.01% (95% Wilson interval 29.81–30.22%), had positive evaluation-period cumulative gross realized PnL.

Primary closed-lot result#

The formation-period 10th-percentile closed-lot win-rate cutoff was 12.20%; the 90th-percentile cutoff was 60.17%.

The observed top-minus-bottom separation was 47.25 percentage points. Cutoff ties were retained, so threshold-defined cohort sizes can slightly exceed one tenth of the eligible sample. This is an association in the eligible, closed-lot sample—not a causal effect and not expected trading return.

Exit-row sensitivity result#

The same study was rerun using formation exit-row win rate while keeping the evaluation outcome and eligibility framework fixed. There is an important production-semantic difference: the per-row realized_pnl_usd field used to classify exit rows deducts allocated entry and exit network fees, while the cumulative full-close lot field used to classify closed lots is gross of those separately stored fees. The sensitivity check therefore changes both accounting grain and fee treatment; it cannot isolate the effect of row-versus-lot grain alone.

The exit-row top-minus-bottom separation was 50.64 percentage points. Formation closed-lot and exit-row rates were strongly correlated (r = 0.892). These results reject an easy marketing story: this simple decile test does not establish that the closed-lot rate predicts the chosen outcome better than the exit-row rate.

What it does establish is that the closed-lot rate retains a strong forward association while measuring the intended unit honestly. Prediction and semantic correctness are separate tests, and a future grain-only comparison should harmonize fee treatment first.

The historical 22.7× wallet observation#

The case that originally motivated the public distinction was recorded during a May 30, 2026 internal endpoint check:

The exit-row figure made the wallet look close to a coin flip. The closed-lot figure said almost none of its fully closed entry lots finished profitably. Because the two production fields also use different network-fee treatment, the 22.7× gap cannot be attributed entirely to exit slicing.

This does not mean the wallet's total PnL was necessarily fake. Win rate and total PnL are different statistics: a wallet can lose frequently and remain net profitable because a small number of winners dominate its losses. It also does not prove that exit-row rates are always higher.

The original response artifact was not retained, so this remains a dated single-wallet aggregate preserved in implementation history. It is not part of the August cohort and is not presented as a competitor comparison.

Why older internal numbers are not headline results here#

Surviving implementation notes referenced an approximately 79,000-wallet field-mapping exercise and a 10–14-percentage-point forward-profitability association. The original cohort queries and result artifacts were not retained. Worse, surviving notes disagree about at least one validation comparator.

Those older numbers explain why Conyr invested in the separate accounting lanes, but they are not publication-grade evidence. This article therefore does not present their MAE values as validated results, does not call the 79K exercise predictive validation, and does not blend the old association with the new temporal study.

One deleted research synthesis did preserve a more specific historical comparison: within an earlier high-past-PnL cohort, 50.9% of the high closed-lot-win-rate group versus 36.6% of the low group was profitable in the next period, a 14.3-point association. Its sample construction, dates, counts, query, and uncertainty are absent, so that result remains provisional history rather than a current claim.

The August study is a new, explicitly frozen analysis under the current post-wipe semantics. It is not described as a reproduction of the older exercise.

From realized rows to one closed-lot outcome#

The production lane is built from Conyr's realized-PnL event stream:

1. An acquisition creates a FIFO entry lot.
2. A disposal consumes one or more lots and emits a realized row for each matched slice.
3. Every positive or negative row contributes to the exit-row counters.
4. Only a row marked as a full lot closure contributes to the closed-lot counters.
5. The full-close row is classified using cumulative gross lot PnL—not only the PnL of the final slice. Separately stored entry and exit network-fee allocations are not deducted from this cumulative production field.
6. Breakevens remain outside both decided-outcome denominators.

The same closed-lot lane supplies related quality inputs: profit factor, a Sharpe-like statistic over closed-lot PnL, winner-versus-loser holding behavior, sizing behavior, loss consistency, ROI percentiles, and runner behavior. Conyr's composite skill_score is an explainable operator-chosen heuristic over normalized inputs. It is not a calibrated probability, credit score, or investment recommendation.

Inspecting both numbers through the API#

The Layer 2 quality endpoint returns both definitions:

curl -s -H "Authorization: Bearer $CONYR_API_KEY" \
  "https://api.conyr.ai/v1/wallet/{address}/quality" \
  | jq '{closed_positions, position_win_rate_pct, raw_exit_win_rate_pct, win_rate_inflation_pts}'

{
  "closed_positions": 184,
  "position_win_rate_pct": 41.2,
  "raw_exit_win_rate_pct": 63.7,
  "win_rate_inflation_pts": 22.5
}

That JSON is an illustrative documented response, not a wallet from either study.

For an integration:

1. Require enough closed-lot evidence; treat rated: false as unknown.
2. Use position_win_rate_pct for fully closed entry-lot outcomes.
3. Keep raw_exit_win_rate_pct when realized-slice behavior matters.
4. Use the gap to identify wallets whose public-looking rate may be driven by exit style.
5. Evaluate profit factor, ROI distribution, loss behavior, automation, provenance, and open inventory alongside any win rate.

One number should never carry the entire verdict.

Limitations#

Closed-only survivorship bias is substantial. Abandoned losing bags, unsellable rugs, open lots, and transfer-only inventory are not marked to market in the chosen outcome. Realized PnL can therefore overstate wallet economics.

The outcome is gross of separately allocated network fees. The cumulative production lot field used for both classification and dollars does not subtract those fee fields. “Positive” in this study is therefore not the same as positive net profit after every recorded cost.

The exit-row sensitivity also changes fee treatment. Production exit rows are classified from a fee-aware per-row field, while closed lots are classified from the fee-gross cumulative field. Their side-by-side separation cannot be attributed to accounting grain alone.

Evaluation eligibility conditions on the future. Requiring at least three decided closed lots in the evaluation window selects wallets that remained active and closed inventory. The rates do not describe every wallet visible in formation.

The study is observational. The decile gap is not a causal effect of win rate, and overlapping wallet behavior or common market regimes can create dependence. Related addresses can violate the nominal independence assumption and make Wilson intervals too narrow.

The corpus is not all-time history. Current wallet state is built from a post-wipe streaming production window. The May case used earlier implementation semantics.

Cost-basis quality still matters. Transferred-in inventory, unsupported paths, missing acquisition history, or incorrect token metadata can leave partial cost basis. Unknown must not be silently converted to zero-cost profit.

Addresses are not people. Explicit bot/arbitrage and malicious-wallet filters do not guarantee that every related, automated, or sybil address is removed.

Positive cumulative gross realized PnL is not a trading strategy result. It is not fee-net or risk-adjusted return, copy-trading performance, latency-adjusted execution, or a guarantee of future profit.

The result is aggregate-public, not row-public. Wallet-level research rows and private exclusion intelligence are not released.

Reproducibility and disclosure#

The aggregate JSON result freezes:

• formation and evaluation windows;
• thresholds and exclusion classes;
• the full logical deduplication identity;
• cohort cutoffs, counts, rates, medians, and Wilson intervals;
• the exit-row sensitivity result;
• analysis reference revision, its scope, and known limitations.

It intentionally omits proprietary query text, internal database and infrastructure identifiers, credentials, raw wallet rows, and private exclusion intelligence. That lets readers audit what was measured without exposing sensitive data shapes or operational details.

The remaining reproducibility boundary is important: independent researchers cannot recreate the exact cohort from the aggregate file alone because the underlying normalized wallet rows are proprietary. The result is aggregate-auditable and method-disclosed, not independently row-reproducible; a rerun requires Conyr's internal data and analysis implementation.

What this changes for wallet intelligence#

The useful conclusion is not that exit-row win rate is worthless. The August sensitivity result shows it can carry information. The conclusion is that the unit must match the question.

If the question is “how often was a realized exit slice profitable?”, count exit rows. If the question is “how often did fully closed FIFO entry lots work?”, count closed lots. If the question is “will this wallet remain profitable?”, neither historical rate is sufficient on its own.

That distinction matters in agent screening, leaderboards, copy-trading research, and smart-money discovery. These systems make wallet-level judgments. They should know whether a reported win is an independently closed lot or one fragment of a larger position before ranking, labeling, or recommending a wallet.

Where to go next#

• Download the frozen aggregate result
• Wallet-intelligence API reference
• Solana wallet-intelligence benchmark
• How to find a profitable Solana wallet
• Public data and methodology
• Coverage and null semantics